fix: surface stateful HTTP session crash cause

[he-yufeng]

Summary

• surface the original stateful session app.run() exception when a POST fails after the session stream closes
• clean up the per-request memory stream on that failure path
• avoid a secondary ClosedResourceError when reporting the failure back to an already closed writer

Fixes #2741.

To verify

• python -m pytest tests\server\test_streamable_http_manager.py -q
• python -m ruff check src\mcp\server\streamable_http.py src\mcp\server\streamable_http_manager.py tests\server\test_streamable_http_manager.py
• python -m ruff format --check src\mcp\server\streamable_http.py src\mcp\server\streamable_http_manager.py tests\server\test_streamable_http_manager.py
• git diff --check -- src\mcp\server\streamable_http.py src\mcp\server\streamable_http_manager.py tests\server\test_streamable_http_manager.py

[he-yufeng]

Rebased this on current main and carried the crash-cause reporting over the updated streamable HTTP code.

Validation run locally on Windows:

python -m py_compile src\mcp\server\streamable_http.py src\mcp\server\streamable_http_manager.py tests\server\test_streamable_http_manager.py
uv run pytest tests/server/test_streamable_http_manager.py -q -k "session_crash_cause or post_error_message" --basetemp .tmp\pytest-2742-20260610 -p no:cacheprovider
# 2 passed, 20 deselected

git diff --check upstream/main..HEAD

The error log lines in that targeted pytest run are expected: the new regression test deliberately crashes the session runner and asserts that the JSON response reports the original session exception instead of only the downstream closed-stream error.