fix: propagate SSE stream errors to waiting requests

[heyhayes]

Summary

Fixes #1401. Also fixes #1789 (closed as duplicate).

When an SSE read timeout occurs during a StreamableHTTP POST request, the pending send_request call hangs indefinitely. The transport catches the exception but never sends an error back through the read stream, leaving the caller blocked on response_stream_reader.receive() with nothing to receive.

This PR fixes error propagation at the transport level so that SSE stream failures produce a JSONRPCError keyed to the original request ID. BaseSession._handle_response routes it to the correct per-request response stream, and send_request surfaces it as MCPError to the caller. This approach keeps failures isolated to the affected request rather than tearing down the entire session.

What changed

_handle_sse_response now sends a JSONRPCError(INTERNAL_ERROR, "SSE stream ended without a response") when the SSE stream ends without delivering a complete response, whether due to a read timeout, network error, or unexpected server close. If a last_event_id was received, reconnection is attempted first; the error is only sent after reconnection is exhausted.

_handle_reconnection returns bool instead of None so callers can distinguish success (response delivered) from failure (attempts exhausted). The method also fixes an infinite recursion bug: the attempt counter was reset to 0 on every stream end (even when no complete response was delivered), which combined with httpx read timeouts causing graceful stream termination meant the reconnection loop could run forever.

handle_get_stream applies the same fix to the GET stream's reconnection loop: the attempt counter only resets when events were actually received during the connection. Empty connections that close immediately count toward MAX_RECONNECTION_ATTEMPTS.

_default_message_handler now logs a warning for exceptions instead of silently discarding them, providing observability for transport errors not tied to a specific request.

Test plan

• New E2E test: client with 0.5s SSE read timeout reads a slow resource (2s server delay), asserts MCPError is raised instead of hanging
• New unit test: _handle_reconnection returns False when called at max attempts
• All 51 existing streamable HTTP tests pass (including reconnection, polling, and multi-reconnection tests)
• All 233 client tests pass
• Linting (ruff check .) and type checking (pyright) clean