#3737)

The MCP session pool manually called transport_ctx.__aenter__() and
session.__aenter__(), attaching anyio cancel scopes to the HTTP request
handler task. When a child task in the transport's TaskGroup failed, it
cancelled the request handler, causing ~20-45% tool call failures.

Changes:
- Run transport/session lifecycle in dedicated asyncio.Task per pooled
  session so cancel scopes bind to the owner task, not request handlers
- Add transport-aware is_closed detection (write stream state, owner
  task health, receive channel count) from PR #3605
- Replace asyncio.wait_for with anyio.fail_after for MCP SDK calls in
  tool_service.py (4 sites), mcp_session_pool.py health checks (4),
  and gateway_service.py explicit health RPC (1)
- Fix release() to clean up dead-owner sessions instead of leaking
  them in _active with consumed semaphore slots
- Fix _create_session() to clean up owner task on CancelledError via
  finally block (not just Exception)
- Add 18 tests for owner task lifecycle, transport-aware is_closed,
  release with dead owner, cancellation cleanup, and prompt pooled
  regression coverage

Tested: 607K+ requests across 200-1000 concurrent users with 0%
pool-related failures. All servers (Fast Test, Fast Time), all
transports (StreamableHTTP, SSE) verified.

Closes #3737

Signed-off-by: Mihai Criveti <crivetimihai@gmail.com>

…ed regression test

The prompt fallback test was only asserting that get_mcp_session_pool()
raises RuntimeError, not that PromptService._fetch_gateway_prompt_result()
actually falls through to the non-pooled sse_client path. Now calls the
real method with mocked pool unavailability and verifies the SSE fallback
session is initialized and get_prompt is invoked.

Signed-off-by: Mihai Criveti <crivetimihai@gmail.com>