Count proxied inference tokens in node metrics

[webdevtodayjason]

Problem

AINode recorded every proxied request with zero tokens because the proxy discarded upstream usage blocks. Streaming replies, Anthropic Messages replies, and embedding requests therefore left token counters and Prometheus token metrics at zero, and the Metrics view could not truthfully show token throughput.

Decision

Read completion or output token totals from non-streamed JSON replies, parse usage from SSE events without changing forwarded bytes, and count output delta events when a stream has no usage event. Record embedding prompt tokens for both fleet and local embedding paths. Restore the token throughput chart as an interval rate derived from the cumulative counter, while retaining the uptime chart.

The harness counter audit found no code change was needed: it differences counters immediately around each task and already states that concurrent node traffic is included.

Proof

• 103 proxy, embeddings, routing, UI, and chart tests passed.
• 168 harness, collector, retention, and Prometheus tests passed.
• ruff check passed.
• No GPU hardware is required because fake engines exercise JSON and fragmented SSE transports end to end.

Changelog

Fixed node token metrics and the Metrics token throughput chart for proxied completion, Messages, Responses, streaming, and embedding requests.

Closes #250

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}