fix(api): return 429 when server is busy

[sufubao]

Summary

• change the default ServerBusyError status from 503 to 429 and label OpenAI-style responses as RateLimitError
• prime streaming responses only in PD-master mode before Starlette sends response headers, so busy failures before the first generated item return HTTP 429 without changing normal-mode header/disconnect behavior
• treat a timeout after streaming has started as an in-flight server_error SSE event, not as a 429/rate-limit rejection
• handle server-busy failures from /get_score
• preserve the Anthropic rate_limit_error envelope for /v1/messages before streaming starts
• log expected pre-stream server-busy control flow at warning level without a traceback

Checks

• python -m pytest -q test/test_api/test_anthropic_extra_body.py (36 passed)
• PD-only stream priming behavior check: normal mode consumes zero items; PD-master replays the prefetched item and propagates a pre-first-item ServerBusyError
• pre/post stream classification check: pre-stream busy propagates for HTTP 429; post-start timeout emits server_error / stream_error
• pre-commit run --files lightllm/server/api_anthropic.py lightllm/server/api_http.py lightllm/server/api_openai.py lightllm/server/api_responses.py lightllm/utils/error_utils.py
• python -m py_compile lightllm/server/api_anthropic.py lightllm/server/api_http.py lightllm/server/api_openai.py lightllm/server/api_responses.py lightllm/utils/error_utils.py