perf(spanner): optimize built-in metrics hot path and harden concurrency

[olavloite]

Optimize hot-path execution in the Spanner client's built-in metrics subsystem and resolve subtle regex boundary and concurrency issues:

• Correctness & Regex Parsing:

  • Add negative lookbehind (?<![a-zA-Z0-9_-]) to GFE and AFE timing regexes to prevent false-positive matches (e.g. safe; dur=55).
  • Guard regex evaluations in extract_front_end_latencies with fast string containment checks ("gfet4t7" in text, "afe" in text).
  • Extract helper _extract_metric_latency and flatten header detection.

• Caching & Allocations:

  • Add bounded LRU cache (maxsize=128) for resource path parsing, yielding a ~9x speedup matching the Java client's caching strategy.
  • Add bounded LRU cache (maxsize=64) for RPC method name formatting.
  • Introduce _ObservableDict to achieve zero-allocation steady-state OpenTelemetry attribute caching with O(1) invalidation on mutation.
  • Consolidate RPC attempt preparation in MetricsInterceptor._prepare_attempt.

• Concurrency & Resource Safety:

  • Extract _BaseAsyncResponseWrapper and synchronize _metrics_recorded via threading.Lock across cancel(), __del__(), and _record_metrics().
  • Safely close unawaited initial metadata coroutines on cancellation.
  • Guarantee ContextVar token reset in MetricsCapture.__exit__ via finally.
  • Harden _safe_decode_utf8 to return "" when passed None.

[gemini-code-assist]

Code Review

This pull request optimizes and hardens the Spanner metrics collection system. Key changes include caching resource path parsing and method name formatting using lru_cache, introducing thread-safe locking and cancellation handling in the response wrappers, caching OpenTelemetry attributes with an _ObservableDict to invalidate the cache on modifications, and refactoring metadata parsing to safely handle various formats. Extensive unit tests have been added to cover these optimizations and edge cases. There are no review comments to address, and I have no feedback to provide.