[mutianf]

Problem

When the vRPC OpenSession API is not served (server returns Unimplemented),
a session dies while still in StateStarting and never activates. onClose's
starting branch bumped the consecutive-failure counter but nothing re-drove
creation: onClosing early-returns for starting handles, and no other of the
sizer's event-driven wake sites fires on a never-activated close.

A cold-start MinSessionCount (5) cohort therefore froze the counter at 5,
never reaching ConsecutiveSessionFailureThreshold (10), so the circuit breaker
never tripped. With nothing re-driving creation and nothing tripping the breaker,
a CheckoutSession caller parked on a deadline-less context hung forever (the
43-min stuck goroutine in the prod round-trip test).

Fix

In onClose's starting branch, after noteAbnormalCloseIfAny, re-drive creation
via spawnTickOnce (gated on sl.ReadyCount() < MaxSessions, mirroring
onClosing's replace-on-death kick). Each re-driven cohort fails again, the
counter climbs to the threshold, the existing breaker trips, and parked
waiters are drained with ErrConsecutiveFailures carrying the underlying
Unimplemented code — only after X consecutive failures, not on the first.
Paced by the throttler's failure penalty, so this is not a tight retry loop.

This replaces the earlier failWaitersOnCreateFailure approach from this PR,
which wrongly failed waiters on the first create failure instead of after a
sustained run. That mechanism (ErrSessionCreateFailed, sessionCreateError,
the Tick call, and the session_pool_waiters_failed_on_create tag) is reverted.

Also narrows the accelerator round-trip test's skipIfSessionUnavailable to skip
on Unimplemented only — the pool now surfaces that code deterministically
instead of hanging, so transient Unavailable / DeadlineExceeded are left to
retryTransient.

Testing

• go test ./internal/transport/ -race — full suite green.
• New tests: TestConsecutiveFailures_StartingDeathsReDriveUntilBreakerTrips
  (end-to-end: starting-deaths re-drive → breaker trips → caller wakes with
  ErrConsecutiveFailures + status.Code == Unimplemented) and
  TestConsecutiveFailures_SingleStartingCloseDoesNotWakeWaiters (a single
  starting-close increments but does not wake waiters).
• Verified against prod cloud-bigtable-spacewalk / mattiefu-test (which
  does not serve OpenSession): the round-trip test now fails fast in 4.72s
  with the breaker trip carrying Unimplemented: Method is not available... and
  cleanly skips — previously it hung for 43 minutes.

[gemini-code-assist]

Code Review

This pull request introduces a mechanism to prevent parked CheckoutSession callers from hanging indefinitely when a session-creation attempt fails and the pool has no remaining path to a ready session. It adds a new ErrSessionCreateFailed error, implements failWaitersOnCreateFailure to wake up stranded waiters, and includes comprehensive unit tests. The reviewer identified a potential data race in failWaitersOnCreateFailure where p.sl.ReadyCount() is called outside of the p.mu lock, and suggested a thread-safe implementation that reads the ready count while holding the lock.