Summary

Google AI returns HTTP 400 / INVALID_ARGUMENT for invalid API keys,
which the shared HTTP classifier treats as an invalid request rather
than an authentication failure.

• Classify these responses as Authentication using the structured
  API_KEY_INVALID reason, with a fallback for Google's known invalid-key
  message.
• Preserve the original HTTP status, provider code, and message.
• Keep unrelated HTTP 400 responses classified as InvalidRequest.

This applies to both model listing and streaming completions through
their shared response classifier.

Fixes DELTA-1C2

Verification

• cargo fmt -p google_ai --check
• git diff --check
• corgi test -p google_ai --lib 'provider_error_tests::' — all 3 tests
  passed.
• Queried Google's live model-listing and streaming-completion endpoints
  with an intentionally invalid key. Both returned HTTP 400,
  INVALID_ARGUMENT, and the structured API_KEY_INVALID reason.
• Passed the actual streaming response body through
  completion_error_from_response in a temporary test and verified it
  classified as Authentication. Removed the temporary test afterward;
  permanent tests cover structured reasons, the message fallback, and
  unrelated invalid requests.

Release Notes:

• Fixed Google AI invalid API keys being classified as invalid requests
  instead of authentication errors.