Shard generated tests across GitHub runners

[nogates]

Summary

• reuse the single compiled Cucumber binary across four GitHub Actions matrix jobs
• assign every scenario deterministically to exactly one shard
• preserve unsharded behavior when the shard environment variables are absent
• revert the temporary src/lib.rs benchmark comments

Baseline

The latest unsharded run executed 1,682 scenarios. Its Run tests step took 5m55s and the complete test job took 6m12s.

Validation

• cargo fmt -- --check
• CARGO_INCREMENTAL=0 cargo check --test main
• CARGO_INCREMENTAL=0 cargo test --test main --no-run
• standalone sharding unit tests (2 passed)
• workflow YAML parse and git diff --check

[nogates]

First CI measurement is green and covers all 1,682 scenarios exactly once:

Shard	Scenarios	Run tests	Complete job
1	394	1m30s	1m45s
2	438	1m31s	1m48s
3	441	1m30s	1m43s
4	409	1m04s	1m17s

The previous unsharded test step/job took 5m55s/6m12s. The slowest shard is now 1m31s/1m48s, reducing the test phase critical path by 4m24s at the step level and 4m24s at the job level.

Despite a slower shared compile in this particular run (4m56s versus 3m49s), the complete workflow improved from 11m10s to 7m34s, a 3m36s wall-clock reduction.