[inductor] Make combo compile-time autotune deterministic-mode safe (#189724)

## Summary

Make combo compile-time autotuning safe under Inductor deterministic mode.

Deterministic mode prohibits the unvetted benchmarking used to tune standalone combo subkernels at compile time. This previously caused compilation to fail through `may_ban_benchmarking`.

When compile-time combo autotuning and deterministic mode are both enabled, this change:

- Skips compile-time subkernel benchmarking.
- Selects per-subkernel block sizes using existing heuristics.
- Disables runtime autotuning for that fallback combo.
- Keeps indirect-indexing nodes eligible for combo fusion because the deterministic fallback performs no synthetic-input benchmarking.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/189724
Approved by: https://github.com/eellison