perf(kv): stage block table with numpy

[adamleex]

Build the paged KV block table through a reused NumPy int32 buffer and a persistent torch view. A microbenchmark on H200 shows lower CPU staging and CPU-to-GPU update time than rebuilding padded Python lists with torch.tensor.

Benchmark

Microbenchmark on H200, measuring CPU staging + GPU copy + synchronize:

Shape	Old Total (µs)	New Total (µs)	Ratio
1x8	20.44	19.79	0.97
1x16	19.84	20.81	1.05
1x32	24.04	21.41	0.89
1x64	27.38	22.47	0.82
2x32	27.23	22.32	0.82
8x64	73.54	45.08	0.61
32x128	452.68	202.28	0.45

[roastduck]

Thank you for your contribution!

But I am wondering why this persistent numpy buffer matters. I thought a malloc of a numpy buffer is much cheaper than the following assignment from block_lists, which involves many list to numpy conversions, and I thought the numpy buffer malloc time can be ignored. What profiling have you done to pinpoint this buffer allocation overhead?