…ptor

Fix an issue where async streaming RPCs were incorrectly wrapped in
_AsyncUnaryResponseWrapper instead of _AsyncStreamingResponseWrapper,
preventing metrics from being properly recorded.

In grpc.aio, streaming call objects (like UnaryStreamCall) implement
__aiter__ (returning an async generator) but do not define __anext__
directly on the call object itself. Because AsyncMetricsInterceptor
checked hasattr(response, "__anext__"), all async streaming calls
(such as ExecuteStreamingSql and StreamingRead) were misclassified as
unary calls and wrapped in _AsyncUnaryResponseWrapper.

Because _AsyncUnaryResponseWrapper only records metrics when awaited
(__await__), and callers iterate over streams chunk-by-chunk rather than
awaiting the stream object itself:
1. Frontend metrics (gfe_latencies, afe_latencies, and missing header
   counts) were never recorded for any async queries or streaming reads.
2. Attempt completion was never recorded when the stream finished. Instead,
   it was only recorded much later when the wrapper was garbage-collected
   (__del__), resulting in wildly inflated attempt latency measurements
   or attempts incorrectly reporting as CANCELLED.

- Explicitly dispatch async interceptor wrappers based on the gRPC call
  type (is_streaming=True for intercept_unary_stream and
  intercept_stream_stream) rather than fragile runtime dunder checks.
- Realign class inheritance so StreamUnaryCall is handled by
  _AsyncUnaryResponseWrapper, adding write() and done_writing() delegation.
- Support sending real HTTP/2 response headers in the in-memory mock Spanner
  server via an add_header() helper, removing monkeypatching of internal
  gRPC classes in integration tests.
- Add unit tests for direct __anext__() consumption (matching
  StreamedResultSet), delegation methods, error handling, and sync/async
  response wrappers.