fix: propagate pre-endpoint errors in sse_client instead of deadlocking

[maxisbey]

Fixes #447

Problem

sse_client launches sse_reader via tg.start(), which blocks until task_status.started() is called. When an error occurs before the endpoint event is received, the except Exception handler tries to send the exception to read_stream_writer. The stream has buffer size 0 and nobody is reading (the caller is still blocked in tg.start()), so send() blocks forever.

This deadlocks in three scenarios:

• Server sends an endpoint event with a mismatched origin (the original report)
• A network error occurs while waiting for the endpoint event
• Server sends a message event before the endpoint event (protocol violation)

Fix

Track whether task_status.started() has fired with a started flag. Before it fires, re-raise exceptions so they propagate through tg.start() to the caller. After it fires, send them to the stream as before — the caller is reading by then.

This generalizes the approach from #975, which fixed the same deadlock but only for SSEError. The dedicated SSEError handler is removed since the flag now covers all pre-endpoint exceptions uniformly.

Behavior change

Previously, SSEError raised mid-stream (after the endpoint was received) would crash the task group. Now it is delivered on the read stream like other post-endpoint errors, letting the session layer handle it. This aligns with how stdio and websocket transports surface mid-stream errors.

Testing

Four new regression tests in tests/shared/test_sse.py cover:

• Endpoint origin mismatch raises ValueError promptly
• Connection error before endpoint raises promptly
• Message-before-endpoint raises RuntimeError promptly
• Post-endpoint errors are delivered via the read stream

All tests use anyio.fail_after(5) guards — if the deadlock regresses, tests fail with timeout rather than hanging CI.

_{AI Disclaimer}