feat(bigtable): add session.Config.EnableDebug to gate sessionz debug state - #20247

… state

The session pool records ~30 sites of debug/snapshot state on every
vRPC and state transition (pick candidates, per-session events/latency
rings, pool-wide latency histograms, per-AFE pick counts, slow-vRPC
log, scaling history, session-lifetime buckets, per-session cluster
counts, time-series sparklines). Even when nobody has wired the
debugview handler and no scraper will ever read the snapshots, the
recorders still allocate on the hot path.

Add a single opt-in session.Config.EnableDebug bool — default false —
that turns off every allocating debug recorder while preserving the
feature-complete debug surface when the flag is on.

Threading:
  session.Config.EnableDebug
    → sessionClient.enableDebug
    → NewSessionPoolImpl(debugEnabled bool)
    → SessionPoolImpl.debugEnabled
    → Session.debugEnabled (via new WithSessionDebugEnabled option,
      wired at createSession time so every session in a pool inherits
      the pool's flag)

Gated recorders (early-return on !p.debugEnabled / !s.debugEnabled):
  - session_pool_debug.go: recordPickDecision, recordLifetime,
    recordTimeSeries, recordSlowVRpc
  - session_pool_scaling.go: recordScaling
  - session_debug.go: Session.recordEvent, Session.recordLatency,
    Session.recordCluster
  - session_pool.go Invoke: pool-wide totalLatencyHist /
    backendLatencyHist / transportLatencyHist bumps
    (RecordTransportOverhead — customer-facing OTel metric — is NOT
    gated; only the debug-only histogram inserts are)

Picker:
  Each of the three AfePicker impls (SimpleAfePicker,
  LeastInFlightAfePicker, LeastLatencyAfePicker) gains a
  recordCandidates bool set at construction. kChoiceMinCost skips the
  per-pick []PickCandidate slice allocation when false; PickDecision
  still populates Winner+Reason for the caller. pickerFromLoadBalancing
  threads p.debugEnabled through so every picker in a pool matches.

Nil-return contract:
  sessionClient.SessionDebug() returns nil when !sc.enableDebug so a
  downstream debugview handler can render a "not enabled" panel
  instead of empty snapshots.

Ungated intentionally (branch check ≥ cost of the work; noted in
session_pool.go comment):
  msgsSent/msgsRecv atomic counters, retries/okRpcs/errorRpcs, and
  the dual-purpose fields (outstanding, remoteAddr, lastConnected).

Tests:
  - New bigtable/internal/transport/session_pool_debug_disabled_test.go
    pins the gate itself: constructs a pool + session with
    debugEnabled=false, drives every gated recorder, asserts every
    ring/map stays at len=0. Without this test a future refactor that
    inverts one of the gates would land green.
  - Existing newTestSession helper enables debug by default (tests
    exercise ring buffers) — WithSessionDebugEnabled godoc explicitly
    calls out the test-author footgun.
  - Existing NewSessionPoolImpl + picker-constructor test callers
    updated (bulk) to pass true so the current debug-on assertions
    still run.

Scope note:
  This PR is scoped to the internal gate mechanics; NewClient does not
  yet accept EnableDebug as a positional argument because it has no
  external caller upstream today (would be a dead knob). Once the
  top-level bigtable.Client wiring lands, that follow-up PR can plumb
  bigtable.ClientConfig.EnableClientDebug into session.Config.EnableDebug
  directly.

Deferred follow-ups (non-blocking, noted per reviewer feedback):
  - Public bigtable.ClientConfig.EnableClientDebug + wiring path
    (needs top-level bigtable.Client session integration first).
  - Debugview "not enabled" banner (debugview package not upstream yet).
  - Compile-time exclusion via //go:build sessionz — build-system work.
  - Sampled snapshots (1-in-N) middle mode.
  - Noop provider vs nil-return refactor.
  - Picker options struct to replace the trailing recordCandidates bool.

Reviewers cleared: session-reviewer, session-component-review,
igor-reviewer (all in parallel per the CLAUDE.md hook).