fix(openai-responses): surface cache read tokens in metadata chunk

[gautham18113]

Summary

Fixes #2407.

OpenAIResponsesModel._format_chunk built the metadata usage dict with only the three scalar token counts (inputTokens, outputTokens, totalTokens). The input_tokens_details.cached_tokens field returned by the Responses API on cache hits was silently dropped, so cacheReadInputTokens was never populated and cache reads were invisible to telemetry and cost tooling.

The fix mirrors what OpenAIModel.format_chunk already does for the Chat Completions path (added in #2115 / #2116): read input_tokens_details.cached_tokens and set cacheReadInputTokens when it is present.

Root cause (before):

case "metadata":
    return {
        "metadata": {
            "usage": {
                "inputTokens": getattr(event["data"], "input_tokens", 0),
                "outputTokens": getattr(event["data"], "output_tokens", 0),
                "totalTokens": getattr(event["data"], "total_tokens", 0),
                # input_tokens_details.cached_tokens dropped here
            },
            ...
        }
    }

After:

case "metadata":
    usage_data: Usage = {
        "inputTokens": getattr(event["data"], "input_tokens", 0),
        "outputTokens": getattr(event["data"], "output_tokens", 0),
        "totalTokens": getattr(event["data"], "total_tokens", 0),
    }
    if token_details := getattr(event["data"], "input_tokens_details", None):
        if cached := getattr(token_details, "cached_tokens", None):
            usage_data["cacheReadInputTokens"] = cached
    return {"metadata": {"usage": usage_data, "metrics": {"latencyMs": 0}}}

Note: cacheWriteInputTokens is not set — the OpenAI Responses API does not return cache-write counts, consistent with the Chat Completions fix in #2115.

Changes

• src/strands/models/openai_responses.py: add Usage import; build usage_data dict first then conditionally populate cacheReadInputTokens
• tests/strands/models/test_openai_responses.py: add two _format_chunk parametrize cases (with and without cache tokens); add test_stream_cache_tokens_propagated and test_stream_no_cache_tokens_when_absent end-to-end streaming tests; add explicit input_tokens_details=None to existing usage mocks so they are not ambiguous

Test plan

• pytest tests/strands/models/test_openai_responses.py — 93 passed, 0 failed
• New test test_stream_cache_tokens_propagated verifies cacheReadInputTokens is set when input_tokens_details.cached_tokens is present on the completed response
• New test test_stream_no_cache_tokens_when_absent verifies cacheReadInputTokens is absent when input_tokens_details is None
• Existing tests updated to set input_tokens_details=None explicitly on usage mocks to prevent false positives