feat(bigtable): add ChainInterceptors and RetryingVRpc for vRPC pipeline - #20185

…/interceptor pipeline

Add two building blocks on top of the vRPC primitives merged in googleapis#20116:

  * ChainInterceptors composes multiple Interceptors around a base Handler
    (first interceptor is outermost, so it sees setup before and teardown
    after the rest).

  * RetryingVRpc is an Interceptor that retries failed virtual RPCs with
    exponential backoff. The default classification is Java-parity:
    Uncommitted attempts always retry, TransportFailure retries only when
    the op is idempotent, and ServerResult retries only if the server
    attached errdetails.RetryInfo. Callers with unusual policies can set
    ShouldRetry to override the default entirely. Per-attempt tracing
    goes through VRpcTracer; the previous attempt's error is tagged onto
    the next attempt's context via WithPrevAttemptErr for downstream
    metrics/logging.

Tests cover chain ordering, success-on-retry, non-retryable and non-gRPC
errors, honoring server RetryDelay (including for DEADLINE_EXCEEDED
carrying RetryInfo), max-attempts exhaustion, and the state-based
classification matrix (Uncommitted always retries; TransportFailure gated
by Idempotent; bare ServerResult never retries by default).

- chain.go: fast paths for len(interceptors) == 0 (no-op wrapper) and
  == 1 (return interceptor as-is), matching gRPC's own
  ChainUnaryInterceptor convention. Avoids the nested-closure alloc on
  every RPC in degenerate cases.

- retrying.go: local attempt counter dropped from atomic.AddInt32 to
  plain int32 (goroutine-local, no synchronization needed). Kept the
  atomic.LoadInt32 on opts.MaxAttempts — that field's doc says
  "Atomic race cap" and external code (config-manager fan-out) may
  swap it while the loop runs.

- retrying.go: swapped time.After for time.NewTimer + explicit Stop()
  on the ctx-cancel exit path so the timer is released immediately
  instead of leaking for the full backoff duration.

Addresses five inline comments on PR googleapis#20185:

1. Strip Java references from doc comments in retrying.go and
   retrying_test.go (nit).

2. Drop InitialBackoff / MaxBackoff / BackoffMultiplier fields from
   RetryingOptions. Backoff is server-driven: honor RetryInfo delay if
   the server sent one, otherwise retry immediately. No client-side
   exponential backoff. Matches Java RetryingVRpc — retryDelay==0 →
   immediate onStateChange(new Idle()).

3. MaxAttempts cap applies ONLY to non-server-directed retries
   (Uncommitted or TransportFailure with no RetryInfo). Server-
   directed retries are uncapped, bounded only by ctx deadline —
   Java parity.

4. Drop the ShouldRetry field. No production caller sets it; the
   field was defensive future-plumbing that never materialized.
   Server RetryInfo already short-circuits the state-based gate;
   the ShouldRetry override was redundant.

5. serverPermitsRetry now gated on RetryDelay.IsValid() — an empty
   or malformed RetryInfo detail is not enough to bypass the
   state-based default. Only a well-formed retryDelay counts as the
   server saying "retry". Matches Java's hasRetryDelay() gate.

Two new tests pin the redesign:
- TestRetryingVRpc_ServerRetryInfoUncapsMaxAttempts — MaxAttempts=3 +
  server RetryInfo → succeeds at attempt 6.
- TestRetryingVRpc_NoClientBackoffOnNonServerRetry — 3 back-to-back
  TransportFailure retries elapse < 50ms (no client-side backoff).

Test coverage now 15 tests (was 14). All pass under -race.

Note: SESSION_SPEC.md line 81 (Client-only inputs table, "3-attempt
cap" row) currently says "hard cap ... overrides any amount of server
RetryInfo" — that text is now stale under the new (Java-parity) code.
The spec file lives on feat/bigtable-sessionz-debug, not on upstream/
main, so the spec edit will land alongside the eventual sessionz-port
of this PR.