feat(schema): track cache creation tokens

[March-77]

What type of PR is this?

feat

Check the PR title.

• This PR title matches the format: <type>(optional scope): <description>.
• The description is user-oriented and clear.
• A separate user-documentation PR is not required; the additive Go API field is documented inline.

(Optional) Translate the PR title into Chinese.

feat(schema): 追踪提示词缓存写入 token

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

en:

Providers can report and bill prompt-cache writes separately from cache reads, but Eino previously exposed only CachedTokens. This made cache-creation volume unavailable to provider integrations and downstream cost accounting.

This PR:

• adds provider-neutral CacheCreationTokens fields to schema.PromptTokenDetails and the model callback DTO;
• preserves cache-creation counts through both message concatenation paths using the existing cumulative-usage maximum semantics;
• adds JSON round-trip and legacy-payload compatibility coverage;
• verifies cache-read and cache-write maxima independently in standard and agentic message tests.

The change is additive. Existing payloads that omit cache_creation_tokens continue to decode with the Go zero value.

Validation:

• go test ./schema ./components/model
• go vet ./...
• golangci-lint run --new-from-rev=origin/main --timeout 5m (0 issues)
• git diff --check

go test ./... was also attempted on Windows. The changed packages pass, while unrelated existing filesystem/path and CRLF-sensitive tests fail under Windows. A WSL rerun could not start because the local WSL instance timed out before go version.

zh(optional):

新增独立的缓存写入 token 统计，并在普通及 Agentic 流式消息合并中保留该累计值，便于准确计算提示词缓存写入成本。

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

Fixes #1150

(optional) The PR that updates user documentation:

N/A

[March-77]

Revalidation update (2026-08-23): this remains mergeable against current main (ebd616c); main is only two unrelated commits ahead, and neither main nor alpha/10 contains an equivalent CacheCreationTokens field. Local validation passes with go test -race ./schema ./components/model.

The Aug 4 scope question on #1150 is valid: this PR establishes the core schema/callback contract, while current eino-ext still stores Claude cache-creation usage in message Extra and its Langfuse/Cozeloop mappings do not consume a cache-write field. I suggest treating provider/observability wiring as explicit eino-ext follow-ups after this core contract lands.

There are still no GitHub Actions workflow runs on this fork PR (only the passing CLA status), so BLOCKED reflects required review, not a failed check. A maintainer review of the core API direction—and workflow approval if the fork banner is shown—would unblock the next step.