feat(eval): token budget tracking and cost reports

[yonatangross]

Summary

REVISED: Descoped from token tracking to duration tracking.

CC Max = flat rate. Token costs irrelevant. Track eval duration instead.

EVAL DURATION REPORT
Skill          Trigger    Quality    Total
commit         45s        120s       2m 45s
assess         38s        95s        2m 13s
api-design     52s        180s       3m 52s

Helps decide whether to run --all or just changed skills.

Depends on #1042, #1043