Rebase/inference usage on meta eval

[kargibora]

Problem

JudgeArena discards token and cost information returned by model providers. Benchmark results therefore do not show the reported usage from generation and judging separately.

Changes

• Collect input, output, total, reasoning, and cached token counts when available.
• Collect provider-reported costs without estimating missing prices.
• Count local vLLM input and output token IDs directly.
• Report totals and coverage by stage and model.
• Save a usage section in run-metadata.v1.json.
• Print per-stage and total usage summaries.
• Preserve existing text/structured outputs and synchronous/asynchronous inference behavior.

How it works

Model response
    -> RequestUsage
    -> active UsageTracker
    -> RunUsage snapshot
    -> console summary and run-metadata.v1.json

The benchmark runner starts one tracking scope. Nested scopes share the same tracker, and the scope is reset when the run finishes or raises an error.

Generation calls use the generation stage and judge calls use judging. Meta-evaluation judges stored completions, so it normally reports only judging usage.

Missing fields remain unknown. Coverage counts distinguish complete, partial, and unavailable usage. Malformed optional metadata does not prevent valid model responses from being returned.

Tracking scope

The tracker records successful responses returned by completed inference calls. These summaries describe observed usage, not a complete billing record. No benchmark scoring policy changes.

Tests

Tests cover canonical and fallback provider metadata, partial reporting, nested tracking, metadata output, failure cleanup, text/structured and sync/async compatibility, and local token counts.

.venv/bin/ruff check judgearena tests
.venv/bin/pytest -q

Result: 395 passed locally.

Raw diff

Relative to #117 at f30aac0. Parent changes are excluded. Pre-existing tests are restored; remaining test reductions apply only to stack-added coverage.

File type	Added	Deleted	Changed
Production Python	438	21	459
Tests	135	1	136
README	4	0	4
Total	577	22	599