feat(agenticopenai): expose prompt cache write usage

[sin-bufan]

What type of PR is this?

feat

Check the PR title.

• This PR title matches the required format.
• The description is user-oriented and clear.
• No separate user documentation update is required; this exposes response billing metadata to downstream integrations.

(Optional) Translate the PR title into Chinese.

feat(agenticopenai): 暴露提示词缓存写入用量

(Optional) More detailed description for this PR(en: English/zh: Chinese).

en:

OpenAI Responses for GPT-5.6 and later reports usage.input_tokens_details.cache_write_tokens separately from cached_tokens because cache writes and reads have different billing rates. agenticopenai currently preserves only cache reads, so downstream observability integrations cannot calculate prompt-cache cost accurately.

This change follows the message.Extra pattern established by #978:

• Exposes positive cache_write_tokens as _eino_openai_cache_write_tokens.
• Adds GetCacheWriteTokens for downstream consumers.
• Covers both Generate and response lifecycle streaming events.
• Keeps cache writes separate from PromptTokenDetails.CachedTokens.
• Reads the field from the OpenAI SDK preserved JSON ExtraFields until the SDK adds a typed property.
• Omits absent, zero, negative, malformed, fractional, and overflowing values.
• Verifies the value survives schema.ConcatAgenticMessages.

Tests:

• go test -gcflags=all=-N\ -l ./...

zh(optional):

OpenAI Responses 的 cache_write_tokens 与 cached_tokens 使用不同计费价格。本改动参考 #978，通过 AgenticMessage.Extra 同时在 Generate 和流式响应中透传缓存写入 token，供 Langfuse 等下游准确计算成本。

(Optional) Which issue(s) this PR fixes:

Related to #978.

[shentongmartin]

Reviewed the implementation and tests. The provider-specific metadata design is consistent with the existing pattern, Generate and Stream paths are covered, and no merge blockers were found.