Add on_poll callback for real-time query monitoring (#723)

[laughingman7743]

WHAT

Add an optional on_poll callback that is invoked once per poll iteration with the current execution object, enabling real-time query monitoring (e.g. live progress in Jupyter). Closes #723.

• New on_poll parameter on BaseCursor.__init__ and Connection / connect(...) (and connection.cursor(...)). Default None → no behavior/perf change when unused.
• Wired at the cursor/connection level only. It rides the existing **kwargs chain into BaseCursor.__init__, so no per-cursor execute() changes are needed — which also avoids adding to the execute() kwargs duplication tracked in Refactor shared execute() kwargs handling across cursor implementations #691.
• Invoked right after each get_query_execution / get_calculation_execution_status, before the terminal-state check, so the final state is delivered too.
• Covers all poll loops, not just BaseCursor.__poll:
  • BaseCursor.__poll — sync + thread-pool async cursors (default/dict/pandas/arrow/polars/s3fs)
  • SparkBaseCursor.__poll — Spark calculations
  • aio native-async loops (AioBaseCursor, AioSparkCursor)
• For Spark the callback receives the per-poll calculation status, so the signature is Callable[[AthenaQueryExecution | AthenaCalculationExecutionStatus], None] (exposed as pyathena.common.OnPollCallback).

Also: centralize on_start_query_execution storage

While adding on_poll to BaseCursor, the sibling on_start_query_execution field was still declared/stored in each of the five synchronous cursors (Cursor, PandasCursor, ArrowCursor, PolarsCursor, S3FSCursor). To keep the two callbacks symmetric, its storage is moved into BaseCursor.__init__ (mirroring on_poll). The per-execute() override and its invocation stay in each synchronous cursor — the broader execute() kwargs consolidation remains #691's scope. Async/aio/Spark cursors inherit the field but do not invoke it (they return the query id immediately); a comment documents this. Behavior is unchanged.

Usage:

def track(execution):
    print(f"State: {execution.state}")

# connection-level default
conn = connect(on_poll=track)
# or per-cursor
cursor = conn.cursor(AsyncPandasCursor, on_poll=track)

WHY

There is currently no public hook into the polling loop, so interactive environments cannot observe query state, elapsed time, or data scanned while a query runs. Requested in #723.

Tests

• Unit (no AWS, deterministic): on_poll fires once per iteration in order incl. the terminal state; None is a no-op; Spark calculation poll variant.
• Integration (cheap SELECT 1): connection-level, cursor-level, and async (AsyncCursor, matching the issue example).
• Existing on_start_query_execution callback tests pass unchanged after the storage move.

make lint (ruff + mypy) and the callback/on_poll tests pass locally.

Refs #691.