ci: run the micro benchmarks (#8142)

Following @HarelM's note after #8122 merged: "I don't know if it's worth adding this to CI, but let's try, worst case we revert." https://github.com/maplibre/maplibre-gl-js/pull/8122#issuecomment-5233617539

One job in the Tests workflow, running `npm run bench` over the 11 `*.bench.ts` files. That is all of it.

It compares nothing. No thresholds, no stored baselines, no uploaded numbers. #982 settled that the only comparison worth trusting is same machine, same session, so a runner has nothing useful to say about a delta. It answers the other question: do the benchmarks still run? Fixtures move, APIs drift, and a bench file can break at runtime where typecheck never looks. The old page failed exactly that way, with three headless runs stalling on network fetches and nothing reporting it.

The suite is node-only and needs no build. The full run takes about 27 seconds here, and on a runner it is likely a couple of minutes once `npm ci` is counted. The 20 minute timeout is for catching a hang rather than for bounding the work.

A module or setup error fails the run loudly, which I checked by breaking a bench file on purpose. A throw inside the measured callback is swallowed by tinybench, so that case shows an empty table instead of a red check.

`npm run bench-e2e` stays out. It needs a dist build and headless Chrome, and its numbers are version comparisons, which is where runner variance misleads most.

This is the last open CI item from #982, though whether that closes the issue is your call. If the job flakes, that is the answer, and the revert is one commit.

- [x] No backports from Mapbox projects
- [x] Related: #982, #8122
- [x] No benchmark scores: CI configuration only
- [x] No CHANGELOG entry, per the CONTRIBUTING rule for benchmark related changes
- [ ] Confirm you have read our AI policy [here](https://github.com/maplibre/maplibre/blob/main/AI_POLICY.md)