fix(spanner): renew channel affinity across retries

[olavloite]

When a dynamic channel pool scales in or drains a channel, read/write transactions must not keep draining channels alive indefinitely across retries, nor should retries remain pinned to a draining channel.

Previously, TransactionRunner::run created a single TransactionAffinity instance outside its retry loop and reused it across all attempts without releasing its read/write channel guard. This caused two problems:

1. If attempt 1 aborted and the channel was marked as draining, retry attempt 2 remained pinned to that draining channel instead of selecting a healthy active channel.
2. The channel's active_rw_transactions guard remained held during backoff sleeps and across attempts, preventing the scaler from closing and sweeping drained channels.

This change resolves both issues:

• Creates a fresh TransactionAffinity per attempt inside the retry loop, matching the behavior of WriteOnlyTransaction and PartitionedDmlTransaction. If a channel started draining during attempt 1, attempt 2 routes to a healthy active channel.
• Explicitly calls affinity.release_rw_guard() as soon as the attempt future completes, ensuring the channel's active R/W counter is decremented before entering aborted backoff sleep.
• Introduces RwGuardState { Active, Released } in TransactionAffinity so guard release is terminal for that attempt, preventing stragglers or late affinity resolutions from resurrecting the guard.
• Adds comprehensive unit tests verifying that retries route to fresh channels, guards are released prior to backoff, draining channels sweep to closed, and dropping the runner future mid-attempt cleans up guards via RAII.

[gemini-code-assist]

Code Review

This pull request updates the Spanner transaction runner to instantiate a fresh TransactionAffinity handle for each transaction attempt instead of reusing a single handle across retries. It introduces a terminal RwGuardState::Released state to prevent late or concurrent operations from re-acquiring a guard on a draining channel. Additionally, the spanner field in DatabaseClient is exposed as pub(crate), and comprehensive tests are added to validate guard release behavior during retries, cancellations, and channel scale-ins. I have no further feedback to provide as no review comments were submitted.

[codecov]

Codecov Report

❌ Patch coverage is 99.44853% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.12%. Comparing base (6ec4f47) to head (be64082).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/spanner/src/transaction_runner.rs	99.37%	3 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##             main    #7002    +/-   ##
========================================
  Coverage   97.11%   97.12%            
========================================
  Files         326      326            
  Lines      114380   114880   +500     
========================================
+ Hits       111085   111575   +490     
- Misses       3295     3305    +10     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.