perf: Optimize CUDA graph batch size selection and padding

[louiswang524]

Improve CUDA graph efficiency by reducing padding waste and using binary search for batch size lookups.

Key improvements:

1. Fine-grained batch size coverage:

   • Small batches (1-7): Capture every size for zero padding waste
   • Medium batches (8-32): Step by 4 instead of 8
   • Large batches (32+): Keep step of 8 for memory efficiency

2. Binary search optimization:

   • Replace linear search with bisect.bisect_left for O(log n) lookup
   • Cleaner code with proper edge case handling

Performance impact:

• Reduces average padding waste from 17.3% to 7.3% (9.9% improvement)
• Particularly beneficial for common small batch sizes (3, 5, 7, 9, 11)
• Trade-off: ~7 additional graphs (~700MB memory for max_bs=160)

Examples:

• Batch size 3: 25% waste -> 0% waste (perfect fit)
• Batch size 5: 37.5% waste -> 0% waste (perfect fit)
• Batch size 9: 43.8% waste -> 25% waste
• Batch size 11: 31.2% waste -> 8.3% waste

The memory overhead is acceptable for modern GPUs (>40GB VRAM) and the improved batch packing efficiency results in better GPU utilization.