fix(responses): preserve cached input token usage

[gruffwy]

Problem

The Responses compatibility layer omits cached input token usage even when providers return it. Streaming providers may expose the value either as top-level cache_read_tokens or as prompt_tokens_details.cached_tokens.

Behavior

• Add standard usage.input_tokens_details.cached_tokens to streaming and non-streaming Responses output when the value is present.
• Preserve explicit zero values.
• Prefer top-level cache_read_tokens, falling back to prompt_tokens_details.cached_tokens only when absent.
• Keep the field absent when cache usage is unknown.
• Do not add cached tokens to total_tokens a second time.
• Do not expose the non-standard cache-write value in the Responses schema.

Tests

• go test ./internal/gateway
• GOFLAGS=-buildvcs=false go test ./internal/app -run 'TestImportConstraints|TestDutyBoundaries' -count=1
• git diff --check