perf: use reduce-overhead compile mode

[vijk777]

Summary

• Added benchmark script for profiling training loop performance
• Found that mode="reduce-overhead" in torch.compile gives 4.9x speedup
• Applied optimization to main training loop

Benchmark Results

Mode	ms/batch	Speedup
none (no compile)	132.95	1.0x
default	98.29	1.4x
default+compiled_backward	92.09	1.4x
reduce-overhead	28.41	4.7x
reduce-overhead+fused	27.25	4.9x
max-autotune	26.21	5.1x

Key Findings

1. reduce-overhead mode is the big win - Uses CUDA graphs to minimize kernel launch overhead
2. Fused Adam provides marginal additional gain (~2%)
3. Compiled backward (autograd) helps ~6% with default mode, but no benefit with reduce-overhead
4. AMP (mixed precision) has compilation bugs with reduce-overhead mode
5. max-autotune is slightly faster but has long compilation time

Expected Impact

Original epoch 3 timing: 195.40s
Expected new timing: ~40s (4.9x faster)