fix(openai): handle server error chunks in streaming responses

[fresh3nough]

Problem

When an OpenAI-compatible server (e.g. llama.cpp) returns an error during streaming, it sends a JSON chunk with an error field instead of the expected choices field. The StreamingChunk deserialization fails with missing field choices, producing a confusing error:

Stream decode error: Failed to parse streaming chunk: missing field choices at line 1 column 99:

This is particularly triggered by subagents/summon, which create concurrent streaming sessions that can overwhelm local LLM servers (regression since the summon extension was introduced).

Fix

• Add #[serde(default)] to StreamingChunk.choices so error-only chunks can be deserialized without crashing
• Add optional error field to StreamingChunk to capture server error responses
• Add check_streaming_error() helper that propagates server errors with clear messages including the original error code, type, and message
• Check for errors in both the main streaming loop and the inner tool-call accumulation loop
• 3 new tests covering: error-only chunks, errors during tool call streaming, and rate limit errors

This follows the same pattern already used by the Google provider (formats/google.rs lines 469-479) which already handles streaming errors correctly.

Testing

• All 31 existing tests in formats::openai::tests pass (no regressions)
• 3 new tests verify error chunk handling:
  • test_streaming_error_chunk_returns_server_error - reproduces the exact error from Stream decode error when using subagents/summon extension (regression in v1.25.0+) #7645
  • test_streaming_error_chunk_during_tool_calls - error mid-tool-call accumulation
  • test_streaming_error_chunk_with_no_choices_no_crash - rate limit errors
• cargo clippy and cargo fmt clean

Fixes #7645
Related: #7364, #7570

[DOsinga]

Thanks for the fix @fresh3nough! Unfortunately this has been superseded by #8031 which landed the same fix (with a slightly different approach — using ProviderError::ServerError so the error is retryable, and handling both the OpenAI and vLLM error formats). Closing this one out, but the contribution is appreciated — it clearly highlighted a real bug that needed fixing!