ci: isolate and cache Numba benchmark smoke

[TATP-233]

Summary

• move the three Numba benchmark contracts from tests/benchmark to benchmark/env
• run benchmark entrypoint smoke and Numba contracts in a dedicated parallel CI job
• persist portable Numba machine-code caches with stable source timestamps
• keep benchmark compilation out of the main coverage job and use CPU-only Torch
• classify the real-spawn dead-collector regression as slow, retaining it in the slow suite without its fixed startup timeout flaking standard CI

Motivation

The Numba benchmark contracts performed cold JIT compilation inside the main coverage job. On GitHub's ephemeral runners, the two largest cases accounted for roughly 131 seconds. Local cold/warm measurements for the extracted contracts were 21.10 seconds and 1.43 seconds respectively.

The real-spawn dead-collector regression also has a fixed five-second child startup wait and has failed intermittently on slow ubuntu-slim runners before reaching the production helper under test. It remains covered by the slow suite.

Validation

• make test-all (completed before PR creation, confirmed by requester)
• make test (1621 passed, 29 skipped, 267 deselected, 1 xfailed)
• uv run --no-sync pytest -q benchmark/env/test_g1_joystick_numba_benchmark.py benchmark/env/test_g1_motion_tracking_numba_benchmark.py benchmark/env/test_numba_random_noise_benchmark.py (13 passed)
• uv run --no-sync python benchmark/smoke_test.py
• uv run pytest -q tests/algos/test_offpolicy_double_buffer_runner.py (66 passed, 1 deselected)
• uv run pytest -q -m slow tests/algos/test_offpolicy_double_buffer_runner.py::test_safe_put_trainer_done_does_not_deadlock_on_real_dead_collector (1 passed)
• CI workflow YAML parse and git diff --check