Add random seed to bench_fwd.py for reproducible benchmarks

[haix716]

Problem

bench_fwd.py generates random inputs on every run with no seed control, making benchmarks non-reproducible across runs. This makes it difficult to compare performance across commits or verify optimization improvements.

Recommendation

Add at the top of the benchmark script:

torch.manual_seed(42)

Impact

Reproducibility. Ensures benchmark results are comparable across runs and commits.