ci: make golangci-lint cache configurable

[trakhimenok]

What changed

• adds a golangci_lint_cache reusable-workflow input, enabled by default
• lets callers disable analysis caching when transfer costs exceed lint time
• exposes the cache invalidation interval, defaulting to the action's seven-day policy
• documents the reusable workflow policy separately from the composite action

Why

Sneat-Go's old lint cache had accumulated analysis for extensions subsequently removed from the repository. A stale-seeded cache made CI slower, but a cache populated exclusively from the decoupled repository is small and materially faster.

Process: faster CI feedback shortens delivery time for fixes to the top-priority SneatBot.

Sneat-Go timing evidence

Scenario	Lint job	Lint-cache details
Cache disabled	2m34s	no restore/save
Stale-prefix first run	6m25s	122 MB restore; 154s save
Stale exact-key warm run	2m49s	106 MB; 50s restore
Truly clean population run	2m36s	no restore; 1.7s save
Clean exact-key warm run 1	1m32s	1.7 MB; 2.0s restore; 11.2s lint
Clean exact-key warm run 2	1m43s	1.7 MB; 1.6s restore; 9.9s lint

The two clean warm runs are 51–62 seconds faster than cache-off. The clean archive is about 98.5% smaller than the stale archive.

Validation

• actionlint .github/workflows/workflow.yml .github/workflows/ci.yml .github/workflows/release.yml
• go test ./...
• go vet ./...
• git diff --check
• shared workflow PR checks, composite-action smoke test, and release-calculator regression checks passed
• Sneat-Go clean population run plus two exact-SHA warm reruns passed

Rollout

After release, delete only Sneat-Go's stale golangci-lint.cache entries, update the Sneat-Go probe to the immutable release commit, and require its normal PR/main/deployment gates.

[trakhimenok]

Controlled Sneat-Go timing result: do not change the default.

The linked Sneat-Go probe #839 compared the current cache-off workflow with a first cache-enabled run and an exact-SHA warm rerun.

• cache-off Lint job: 2m34s
• first cache-enabled Lint job: 6m25s
• exact-key warm Lint job: 2m49s
• warm cache restore: 106 MB in 50.3s
• warm actual lint analysis: 35.9s
• first-run cache save: 154.0s

Although the warm analysis itself was faster, cache transfer erased the gain. The exact warm run was still 15s slower overall, while a changed key imposed several minutes of overhead.

Closing without merge; v1.10.4 remains the active released policy with skip-cache: true.

[trakhimenok]

Reopening for a stricter follow-up requested after reviewing the first probe.

The first experiment restored a prior seven-day cache through the action's prefix fallback, so it did not answer whether a cache built only from the decoupled repository is worthwhile. This branch now exposes golangci_lint_cache_invalidation_interval; the Sneat-Go probe sets it to 1, a namespace not present in the repository's current cache inventory.

The follow-up will measure:

1. a guaranteed no-restore population run;
2. an exact-SHA rerun of the newly populated clean cache.

No existing cache is deleted.