perf(spanner): make async query execution event-driven - #14933

Asynchronous queries share a bounded background thread pool to read rows from
query streams and dispatch them to application callbacks. Previously, each
active query held onto a background thread for its entire lifetime, parking
that thread whenever the query was waiting for more data from the server,
waiting for a full row buffer to be drained, or paused by the application
callback. Running many concurrent or paused queries could exhaust the shared
thread pool and stall other queries waiting for an available worker thread.

This change switches asynchronous query execution to an event-driven model so
background threads are only used while actively moving buffered rows:

- Event-driven row production: Background tasks read rows only while stream
  data is already buffered locally and space remains in the row queue,
  releasing the worker thread immediately whenever a query is waiting on the
  network, the row buffer is full, or consumption is paused. Incoming stream
  data, buffer consumption, and resume calls re-schedule row production as
  needed.
- Serialized stream startup and consumption: Completes stream initialization
  before row production begins so underlying query iterators are never
  accessed concurrently across threads, while preserving startup error
  translation for queries that inline transaction creation.
- Prompt cancellation and clean lifecycle termination: Delivers cancellation
  directly to paused or in-flight callbacks without requiring a separate
  resume call, closes underlying streams outside internal locks, and waits for
  all active tasks to finish before completing the result future and releasing
  executor resources.
- Strict callback thread scoping and interrupt handling: Restricts cursor
  advancement and row reads in asynchronous mode to the active callback
  invocation thread, and cancels the query while preserving the thread
  interrupt status if a blocking list conversion is interrupted.