fix(anthropic): require message_stop before finishing stream

[ethanndickson]

Context

#232 added fantasy.NewIncompleteStreamError() and gated the Anthropic post-loop on acc.StopReason == "". That catches the user-facing-content cases: a stream that closes anywhere from the first content_block_delta up to (but not including) the message_delta event leaves acc.StopReason empty and is correctly reported as a retryable incomplete stream.

What that check still admits is the narrow window between message_delta and message_stop. message_delta carries the stop_reason and sets acc.StopReason non-empty, so the existing guard passes; but the protocol-required message_stop terminator never arrives. The stream then emits StreamPartTypeFinish with a real-looking finish reason, hiding the fact that the transport closed unexpectedly in a window the Anthropic SSE protocol documents as impossible on success.

Fix

Require both terminal signals before yielding Finish:

• A new sawMessageStop bool is flipped to true inside the chunk-type switch when a message_stop event arrives.
• After the loop, if sawMessageStop is false or acc.StopReason is empty, the stream yields fantasy.NewIncompleteStreamError() (retryable *fantasy.ProviderError with Cause: io.ErrUnexpectedEOF) instead of Finish. ctx.Err() takes precedence so caller-initiated cancellations stay distinguishable from synthetic transport errors — a small attribution improvement on the existing path as well.

This matches the pattern Anthropic ships in their own newest SDK: MessageAccumulator in anthropic-sdk-java (PR anthropics/anthropic-sdk-java#178) raises IllegalStateException("'message_stop' event not yet received.") on MessageAccumulator.message() and has dedicated unit tests for both messageNotStarted and messageNotStopped.

Tests

A new table-driven test TestStream_RequiresMessageStopBeforeFinish covers:

• complete happy path (both events present) → Finish
• message_delta with stop_reason present but message_stop missing → retryable incomplete-stream error (the new window this PR closes)
• message_stop present but stop_reason missing → retryable
• a server-sent error SSE event → surfaced verbatim and not retryable

The pre-existing TestStream_TruncatedWithoutStopReason (from #232) covers the simpler case where neither terminal event arrives and is unchanged.

Follow-ups

Similar terminal-event gaps exist in a few of the other providers in this repo; those will land in separate PRs so each can be reviewed in isolation.