Fix Triton pointer offsets for large batches

[HuXinjing]

Summary

This fixes large-batch pointer arithmetic in the V2 Triton MaxSim kernels.

For large document batches, pointer offsets such as:

D_ptr + batch_idx * Nd * d + ...

can cross the signed int32 element-offset boundary before GPU memory is exhausted.
With a real ColQwen page shape of Nd=747, d=320, the limit is about:

floor((2^31 - 1) / (747 * 320)) + 1 = 8984 documents

That caused flash_maxsim_v2_multiquery to fail with CUDA illegal memory access
around 8990 documents even though GPU memory usage was only about 4.5 GiB.

Changes

• Cast Triton program ids and index vectors used in pointer arithmetic to tl.int64.
• Apply the fix consistently to:
  • _flash_maxsim_v2_kernel
  • _flash_maxsim_v2_multiquery_kernel
  • _flash_maxsim_v2_mq_dimtile_kernel

Validation

Environment:

• GPU: RTX 4090 24GB
• Query shape: 44 x 320
• Document/page shape: 747 x 320
• dtype: fp16
• Kernel: flash_maxsim_v2_multiquery

Before this change:

Documents	Result
8950	ok
8990	CUDA illegal memory access

After this change:

Documents	Result	Peak GPU memory	Mean scoring latency
8990	ok	4559 MiB	9.20 ms
10000	ok	5019 MiB	10.22 ms
16384	ok	7929 MiB	16.56 ms
32768	ok	15399 MiB	32.86 ms
49152	ok	22869 MiB	49.14 ms
53248	expected CUDA OOM during tensor allocation	-	-

The failure mode after the fix is normal GPU memory exhaustion rather than an
illegal memory access.