[codex] Re-land remote cold start workflow headers

[hansent]

Summary

• re-land remote cold start model metadata aggregation in workflow response headers
• restore the dedicated request_metrics module and its focused tests
• fix the stale http_api.py middleware copy by using the shared request_metrics implementation

Root Cause

PR #2209 merged into main and was then reverted in #2222 because inference/core/interfaces/http/http_api.py still contained an older local copy of GCPServerlessMiddleware after the shared request_metrics.py extraction. That stale copy referenced symbols that were no longer imported there, which triggered the CI flake8 F821 undefined name failures.

Why

Workflow responses should include remote model IDs, cold start counts, load times, and detailed load metadata when remote execution happens. Without this reland, top-level workflow headers only reflect local execution and under-report remote cold starts.

Validation

• python3.10 -m py_compile inference/core/interfaces/http/http_api.py inference/core/interfaces/http/request_metrics.py inference_sdk/http/utils/executors.py inference_sdk/config.py inference/core/managers/model_load_collector.py
• /Users/hansent/.venv/bin/python -m pytest tests/inference_sdk/unit_tests/test_config.py tests/inference_sdk/unit_tests/http/utils/test_remote_processing_time_collection.py tests/inference/unit_tests/core/interfaces/http/test_remote_processing_time_middleware.py tests/inference/unit_tests/core/interfaces/http/test_model_response_headers.py

Context

This PR re-lands the intended behavior from #2209 after the revert in #2222, with the missing import/middleware cleanup included.