fix(mistralai): standardize API error handling

[jeanprbt]

Summary

Standardize Mistral error handling across LLM, STT, and TTS while preserving provider error details and retry semantics.

Motivation

With an invalid LLM model such as foo, Mistral returns HTTP 400 with error code 3003.

Model foo is currently not in use.

The LLM plugin previously handled Mistral's SDKError as a generic exception, turning this non-retryable API response into a retryable APIConnectionError. One user turn therefore produced four misleading Connection error events, the
initial request plus three retries.

Changes

• Translate Mistral SDKError and HTTPValidationError into APIStatusError across LLM, STT, and TTS.
• Preserve the provider status, body, and x-request-id.
• Translate actual asyncio/httpx timeouts into APITimeoutError.
• Preserve LLM stream retryability so errors after output begins are not retried, avoiding duplicated partial output.

Ordinary 4xx responses are now non-retryable, while transient statuses continue to follow APIConnectOptions.

Tests

• HTTP 400 is attempted once and remains an APIStatusError.
• HTTP 422 preserves its status, body, and request ID.
• An error after the first streamed chunk is not retried.

uv run pytest tests/test_plugin_mistralai_llm.py -q

[devin-ai-integration]

Devin Review found 1 potential issue.

1 flag not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)