feat: integrate CozeLoop tracing and prompt policy

[YangYuS8]

What changed

• initialize the official Eino CozeLoop callback adapter only when explicitly enabled
• use a strict allowlist parser that removes raw model input/output, Provider payloads, authorization values, and unrestricted message content from traces
• create one safe root span per Agent run with request, conversation, tenant, release, Prompt version, scope count, and Tool count metadata
• add a CozeLoop PromptHub-backed policy appendix with version tracking, bounded refresh, last-approved caching, and bundled fallback
• keep the compiled Nivora safety rules authoritative even when a remote Prompt is active
• expose active Prompt source/version through /version
• degrade to no-op tracing and bundled policy when CozeLoop initialization or refresh fails
• add tests for trace redaction, Prompt versioning, fallback behavior, and unsafe Prompt rejection

Production behavior

CozeLoop is optional. Trace export failures and PromptHub outages do not fail customer conversations. Remote Prompt content is accepted only from system-role messages, is size-limited, and cannot override the compiled safety policy.

Locked dependencies

• github.com/cloudwego/eino-ext/callbacks/cozeloop v0.3.1
• github.com/coze-dev/cozeloop-go v0.1.22

Validation

The current head passes the final read-only workflow:

• clean go mod tidy
• gofmt
• go vet ./...
• go test -race ./...
• go build -trimpath ./...