Add token usage reporter and optimization skill

[V3RON]

Why this change is needed

Skillgym needs a built-in way for agents to compare billable token usage without parsing the full suite result or relying on ad-hoc artifacts. This adds a strict token-usage reporter for optimization loops and a bundled token-optimization skill that tells agents how to reduce token cost safely while keeping benchmark behavior intact.

For Skillgym users, this means token-cost tuning can stay inside the normal skillgym run workflow: run a baseline, make one small change, compare compact JSON output, and fall back to the existing artifacts when a row fails.

Closes #43.

Architecture and implementation details

• add a built-in token-usage reporter and wire it through reporter loading, CLI help text, and the public reporter exports
• keep reporter stdout to one compact JSON object with top-level passed, billable, artifacts, and rows fields so agents can parse it reliably
• compute per-row billable only for passed rows with provider-backed normalized totals; keep failed, derived, and unavailable rows in the output with billable: null
• aggregate top-level billable from comparable passed rows only, including repeated executions via successfulRepetitions
• avoid a second token artifact format and reuse the existing suite-run artifact directory for debugging when rows fail
• add the bundled token-optimization skill and document the required workflow: explicit target, minimal protecting suite, passing baseline first, one small edit at a time, before/after comparison, and snapshots only as optional follow-up protection
• update the bundled skill docs, reporter docs, CLI skill coverage, and DICTIONARY.md with the approved token-usage and token-optimization terms
• include the small fixture repairs needed to keep the requested local CI gates green while landing the reporter slice

Validation

• pnpm run fmt:check
• pnpm run test
• pnpm run typecheck
• pnpm run build
• ran the full check set before each commit