fix(openai): report prompt cache write tokens in usage

[1stRunciter]

Current behavior

OpenAI Responses cache writes remain inside InputTokens, while CacheCreationTokens stays zero.

Expected behavior

Subtract provider-reported cached reads and cache writes from normalized input tokens, and expose writes through the existing CacheCreationTokens field. Normalized input is clamped at zero, while TotalTokens retains the existing raw input-plus-output aggregate.

Compatibility

Absent or zero cache-write values preserve existing behavior. This makes no public API changes.

Path coverage

The shared mapping covers generate, response.completed and response.incomplete streams, successful object generation, both object-stream terminal events, and no-text object-generation errors.

Evidence

OpenAI documents cache_write_tokens and cached_tokens under Responses usage.input_tokens_details in the prompt caching guide.

Case	input_tokens	cached_tokens	cache_write_tokens
cold write	3438	0	3435
warm read	3438	3423	12
below cache minimum	14	0	0

Scope

Responses only. Chat Completions and proxy-provider behavior remain unchanged.

Validation

• gofumpt -w providers/openai/responses_language_model.go providers/openai/openai_test.go
• golangci-lint run
• go test ./... -count=1 -timeout=30m