[ci] Shard build workers by measured bundle build times (#37353)

`build_and_lint` assigns its `[bundle, bundleType]` pairs to 25 workers
per channel by round-robin, which leaves the slowest worker with 56-62
seconds of rollup time while the mean worker has 40-42 seconds (measured
from the timestamped `BUILDING`/`COMPLETE` lines in recent `main` run
logs).

This change shards by measured build time instead, and the measurement
maintains itself:

`yarn build` writes the timing results into `build/__shard_timings__/`,
which rides along inside the existing per-worker artifacts.

`process_artifacts_combined`, which already downloads all 50 artifacts
and is off the critical path, combines them into `build-weights.json`
and saves it to the actions cache under a per-run key. Readers restore
the most recent entry via a `restore-keys` prefix. Only pushes can save
to cache keys that PRs can read, so pull request runs benefit from the
weights but cannot poison them.

The new measurement is always written verbatim rather than merged with
previous weights, so removed bundles drop out instead of accumulating.

A per-bundle diff against the previous weights is logged so that we can
monitor whether single-run variance is too high, in which case shards
should be determined from timings across the last N runs instead. Even
on this PR a perfect prediction would've only gained us ?s for the
slowest shard.



Co-authored-by: Claude Code (kimi-k3[1m]) <noreply@anthropic.com>