[shuaiyafei]

Summary

Fixes #773.

Declare prompt_eval_cached_count on BaseGenerateResponse so GenerateResponse and ChatResponse preserve the cache metric already returned by the server. It defaults to None when absent; an explicit zero is preserved.

The field matches the server's Metrics.PromptEvalCachedCount and the generate/chat response definitions in docs/openapi.yaml. This changes no request parameters or client control flow.

Validation

• Added 14 regression cases: model parsing/serialization with an absent, zero, or nonzero count; synchronous and asynchronous generate/chat calls; ordinary responses and terminal NDJSON chunks.
• All 14 fail on the unchanged implementation because the field is missing.
• python -m pytest ollama tests -q: 136 passed.
• ruff format --check .: 43 files already formatted.
• ruff check . and git diff --check: passed.

Client tests use a local HTTP test server and the real SDK request/response parsing. No live model inference was run. Environment: Python 3.12.15, locked runtime dependencies.

AI assistance: Codex investigated the issue, implemented the change, and ran the checks above.