fix(openai): preserve DeepSeek cached token usage

[m11y]

Summary

• map the DeepSeek OpenAI-compatible prompt_cache_hit_tokens usage extension into Eino PromptTokenDetails.CachedTokens
• preserve a non-zero standard OpenAI prompt_tokens_details.cached_tokens value, while falling back to the DeepSeek extension when a gateway supplies an empty standard details object
• fall back to zero for missing or malformed extensions

Problem

DeepSeek returns cache hits as a top-level usage field. The underlying go-openai client preserves unknown usage fields in Usage.ExtraFields, but toEinoTokenUsage only reads the standard nested OpenAI field. As a result, both Generate and Stream responses lose the cache-hit count while being converted to Eino metadata. Some compatible gateways may also synthesize prompt_tokens_details with a zero cached value, so zero must not mask a non-zero DeepSeek extension.

Tests

• GOTOOLCHAIN=go1.24.6 go test -run TestToEinoTokenUsageCachedTokens .
• GOTOOLCHAIN=go1.24.6 go test -gcflags="all=-N -l" ./...
• GOTOOLCHAIN=go1.24.6 go vet ./...