feat(bigtable): add ClientConfigurationManager - #19986

ClientConfigurationManager polls GetClientConfiguration on a fixed
interval, parses the response into a typed clientConfig, and fans
changes out to registered listeners. Supports RPC retry with capped
randomized exponential backoff, a validity-window fallback to defaults
on prolonged outage, and Close() shutdown that waits for in-flight
polls so listener callbacks cannot fire against state that is about to
be torn down.

Listener delivery is monotonic-by-sequence (CAS-guarded against the
register-then-fire race with concurrent polls). Two cheap diff layers
suppress redundant work: a whole-config == check in poll() skips the
fan-out entirely when the polled config is unchanged, and the typed
AddSessionPoolListener / AddSessionLoadListener wrappers each hold the
last value they delivered and short-circuit when their own slice is
unchanged — mirroring Java's ListenerEntry.maybeNotify pattern.

Wired into Client behind a new ClientConfig.EnableSession flag (off by
default). When set, NewClientWithConfig constructs the manager using
the data-plane BigtableClient + instance-scoped request metadata and
Start()s it; Close() stops the manager first so it cannot fire
listeners against a tearing-down pool.

The session-pool ChannelPrimer / DirectAccessChecker / pool-factory
wiring that consumes this manager is intentionally deferred to a
follow-up PR for review scope.

…backoff

Address review on googleapis#19986:

- Per-field doc comments on pollingConfig, sessionConfig,
  channelPoolConfig, sessionPoolConfig, and loadBalancingOptions so the
  config surface is self-describing rather than implied by the proto
  field names.
- Expand the configListener type doc to specify what seq carries
  (the manager's monotonic configSeq at fire time) and how listeners
  can rely on monotonicListener's guard against reordered deliveries.
- Move the defaultClientConfig literal into its own
  default_client_config.go so it reads as a tunable defaults table
  separate from the polling machinery.
- Hoist the 5s per-RPC poll timeout into a named pollDeadline const
  reused by both Start()'s eager initial poll and pollingLoop()'s
  periodic polls, instead of two open-coded 5*time.Second literals.
- Collapse the retry-backoff cap to a one-liner using built-in min:
  min(1<<min(i, 30), maxBackoffSeconds). Preserves the overflow guard
  (shift width capped at 30) and the absolute backoff cap, and drops
  the explicit <=0 fallback branch that's no longer reachable.

ClientConfiguration.Polling is a oneof; until now the manager only read
the PollingConfiguration case and silently ignored StopPolling. The
StopPolling bool exists as a server-side backstop against excessive
GetClientConfiguration RPCs, so a client that ignores it forces the
control plane to choose between living with the load or returning
errors instead.

Carry StopPolling through as a typed field on pollingConfig, set it in
parseConfig from the oneof, and exit pollingLoop at the top of the
iteration after a poll that delivers it. The current configuration
stays in effect — no fallback to defaults — so the rest of the client
keeps using the most recent server-honored config; only further polls
are suppressed. Close() still tears the manager down normally.

…ation

Address mutianf's review on googleapis#19986: "why do we need to parse the config
to go type instead of using the proto object directly?"

Drop the parallel Go-struct hierarchy (clientConfig, pollingConfig,
sessionConfig, channelPoolConfig, sessionPoolConfig, loadBalancingOptions,
channelPoolMode + the Strategy*/mode* constants) and the parseConfig /
parsePollingConfig / parseSessionConfig / parseChannelPoolConfig /
parseSessionPoolConfig pipeline. ClientConfigurationManager now stores
currentConfig, defaultConfig, and lastResponse as
*bigtablepb.ClientConfiguration and forwards the raw proto to listeners,
so nothing on the read path has to reason about two shapes of the same
data.

Field-level fallback is done at the accessor site by three small
helpers — pollingInterval, validityDuration, maxRPCRetryCount — that
read from the proto and reach into defaultClientConfig when the server
left the enclosing message out of the oneof / left the field nil.
StopPolling and SessionLoad are cheap enough to read via generated
getters (cfg.GetStopPolling() / cfg.GetSessionConfiguration().
GetSessionLoad()) so no wrappers for those. Change detection swaps ==
for proto.Equal; getConfig() and listener fires now return proto.Clone
copies so consumers can't race the manager on mutation.

Also addressed in the same pass:
- The two open-coded time.Hour * 24 * 365 * 100 sentinels (validUntil
  init in NewClientConfigurationManager, validity-window reset in
  poll()'s fallback branch) are now derived from
  validityDuration(defaultClientConfig). One source of truth in the
  defaults table, no magic literal in the polling code.
- AddSessionPoolListener no longer round-trips through a Go struct that
  dropped Headroom / NewSessionCreationBudget / LoadBalancingOptions /
  NewSessionCreationPenalty / ConsecutiveSessionFailureThreshold. It
  now forwards the server's SessionPoolConfiguration verbatim, so
  downstream PoolSizer / SessionPoolImpl.UpdateConfig see every field
  the control plane sent.
- Additive observability from feat/bigtable-sessionz-debug: PollEvent
  ring buffer (maxPollHistory=100), lastResponse / lastFetchedAt /
  lastErr / lastErrAt capture, Snapshot() returning ConfigSnapshot for
  the configz debug UI. sessionz-only debug-tag calls are skipped —
  their infra doesn't exist on this branch.

defaultClientConfig is now a *bigtablepb.ClientConfiguration proto
literal (300s poll / 100y validity / 5 retries / SessionLoad=0 / 2-25
servers / 10 sessions per server / DirectAccessWithFallback with 60s
check + 0.8 error threshold / LeastInFlight LB). Values unchanged from
the previous Go-struct table.

Deferred to follow-up PRs (not addressed here):
- mutianf's "pass in a channel instead of BigtableClient" refactor —
  API surface change.
- mutianf's "we might have a server-side version number; leave your
  monotonic seq out" — semantic listener-contract change.
- TelemetryConfiguration handling — no consumer on this branch yet;
  raw proto passively preserves it in currentConfig for later.