ci: skip gh-pages deploy for partial benchmark runs

[lukekarrys]

Filtered workflow_dispatch runs (e.g. only registry-clean/registry-lockfile) are the fastest way to validate registry changes, but deploying their results breaks gh-pages: the day's data is missing most variants and charts don't render. This makes partial runs safe to fire at will.

A Detect partial run step in detect-changes marks any dispatch whose inputs differ from the defaults (fixtures, variations, binaries, warmup, runs) as partial, and the deploy job now skips when partial_run == 'true'. Scheduled and push runs are unaffected. Detection is automatic rather than an opt-in input so a filtered run can never pollute gh-pages by accident; there is deliberately no override.

The process job also gains a Summarize results in log step that prints every fixture/variation's per-command mean and raw times into the job log and $GITHUB_STEP_SUMMARY, so partial runs are fully readable from the Actions UI without downloading artifacts (raw results-* artifacts still upload from each matrix job as before, 7-day retention).

Second commit fixes two pre-existing quirks found while making this change. The workflow_dispatch variations default included build/build-cache, which a defaults dispatch crossed with every fixture (build×astro, etc.) while scheduled runs only add them for next via include: pairs — the dispatch default now matches the scheduled fallback so a defaults dispatch is identical to a scheduled run, and the include pairs still supply build/build-cache/run. And the single concurrency group meant a dispatch during the daily window silently cancelled the in-progress scheduled run, losing that day's data; the group is now scoped by event_name so dispatches only cancel other dispatches, pushes cancel pushes, and scheduled runs are never cancelled by manual activity.