Insert prompt cache tokens in ChatDatabricks response metadata

[viadark]

Summary

• Databricks Model Serving returns prompt caching token info (Anthropic: cache_creation_input_tokens, cache_read_input_tokens / OpenAI: prompt_tokens_details.cached_tokens) in CompletionUsage,
  but ChatDatabricks was discarding them from llm_output and streaming UsageMetadata
• Extract and insert cache tokens in both invoke (_convert_response_to_chat_result) and stream (_extract_completion_usage_from_chunk, _build_usage_chunk_from_completions) paths

Changes

chat_models.py

Method	Change
_convert_response_to_chat_result	Add Anthropic/OpenAI cache tokens to llm_output["usage"] and top-level llm_output
_extract_completion_usage_from_chunk	Include cache tokens in dict fallback path
_build_usage_chunk_from_completions	Build InputTokenDetails with cache info in dict fallback path

test_chat_models.py

Test	Description
test_convert_response_to_chat_result_anthropic_cache_tokens	Verify Anthropic cache tokens in llm_output
test_convert_response_to_chat_result_openai_cache_tokens	Verify OpenAI cached_tokens in llm_output
test_convert_response_to_chat_result_no_cache_tokens	Verify no cache keys when absent
test_chat_databricks_stream_with_claude_cache_tokens	Verify Claude cache tokens in streaming UsageMetadata