refactor: subscribe metrics plugins to generation_batch_* hooks #1182

Background

PR #1181 (Phase 1 of #444) introduces three new hooks for raw-path generations:

• GENERATION_BATCH_PRE_CALL
• GENERATION_BATCH_POST_CALL
• GENERATION_BATCH_ERROR

The tracing plugin subscribes to all three so that generate_from_raw emits one span per call (per OTel GenAI semconv). The metrics plugins do not — mellea/telemetry/metrics_plugins.py only subscribes to per-MOT hooks:

As a result, generate_from_raw calls produce zero metrics today. Tracing covers it; metrics don't.

Scope

Pure additive. For each metrics plugin, add a parallel hook handler that subscribes to the corresponding generation_batch_* hook and records the same metrics from the batch payload. Estimated ~30–40 LOC.

Acceptance criteria

• All four metrics plugins subscribe to the matching generation_batch_* hook.
• Token, latency, cost, and error metrics are recorded for generate_from_raw calls in addition to generate_from_chat_context.
• Test added that calls generate_from_raw with metrics enabled and asserts each counter/histogram is incremented.

Reference

PR #1181 (the tracing plugin in mellea/telemetry/tracing_plugins.py shows the batch-hook subscription shape to mirror).