Summary

Fixes #1401.

The default _default_message_handler in ClientSession silently discarded all messages — including exceptions. When a transport error (e.g. SSE read timeout) was delivered through the read stream, the handler did nothing but await anyio.lowlevel.checkpoint(), making it impossible to diagnose why requests were hanging.

This change adds a logger.warning() call with full exception traceback (exc_info) when the default handler receives an Exception. This provides immediate observability for transport errors without breaking backward compatibility — callers who want custom behavior (re-raising, suppression, etc.) can still pass their own message_handler.

What changed

• _default_message_handler now logs a warning with the exception traceback when it receives an Exception, instead of silently discarding it
• Added test verifying the default handler logs exceptions with proper exc_info
• Added test verifying custom handlers can still suppress exceptions if desired

Why not re-raise?

Re-raising in the default handler would terminate the receive loop and send CONNECTION_CLOSED errors to all pending requests. While this prevents hangs, it also affects non-critical exceptions routed through _handle_incoming (e.g. responses with unknown IDs). Logging provides observability without this side effect. Transport-level error propagation to response streams is addressed separately in #2122.

Test plan

• New test: test_default_message_handler_logs_exceptions — verifies the exception is logged at WARNING level with correct exc_info
• New test: test_custom_message_handler_can_suppress_exceptions — verifies custom handlers can still suppress exceptions
• All 13 client session tests pass
• All 8 shared session tests pass
• ruff check clean
• pyright clean