Fix apply_bit_mask cuda implementation

[Jialin]

Issue

Originally, most of the apply_token_bitmask unit tests are failed due to cuda invalid memory access.

pytest tests/python/test_token_bitmask_operations.py 

Fix

With some debugging, we found that the root cause is due to putting logits.stride and bitmasks.stride in the wrong order.

Benchmark

With the fix, we also loop in cuda comparison in the benchmark runs, which still show significant improvements on top of triton implementation.

|   Batch |   Vocab |   Masked cnt |   Torch Compile |         Triton  |           CUDA  |
|    size |    size |              |     Baseline us |    us (speedup) |    us (speedup) |
|--------:|--------:|-------------:|----------------:|----------------:|----------------:|
|       1 |  128000 |            1 |            5.85 |    5.41 (1.08x) |    5.46 (1.07x) |
|       1 |  128000 |        64000 |            5.84 |    6.01 (0.97x) |    6.24 (0.94x) |
|       1 |  128000 |       127000 |            5.84 |    6.09 (0.96x) |    5.95 (0.98x) |
|       8 |  128000 |            1 |           10.75 |    5.86 (1.83x) |    5.90 (1.82x) |
|       8 |  128000 |        64000 |           10.75 |    7.59 (1.42x) |    9.85 (1.09x) |
|       8 |  128000 |       127000 |           10.77 |    7.85 (1.37x) |    8.06 (1.34x) |
|      64 |  128000 |            1 |           48.59 |   13.10 (3.71x) |    9.68 (5.02x) |
|      64 |  128000 |        64000 |           48.59 |   45.43 (1.07x) |   38.76 (1.25x) |
|      64 |  128000 |       127000 |           48.58 |   32.84 (1.48x) |   26.29 (1.85x) |
|     512 |  128000 |            1 |          349.84 |   67.34 (5.20x) |   37.06 (9.44x) |
|     512 |  128000 |        64000 |          346.94 |  330.36 (1.05x) |  256.53 (1.35x) |
|     512 |  128000 |       127000 |          345.54 |  249.66 (1.38x) |  157.51 (2.19x) |
|    4096 |  128000 |            1 |         2895.83 |  494.47 (5.86x) | 249.96 (11.59x) |
|    4096 |  128000 |        64000 |         2863.31 | 2517.85 (1.14x) | 1993.29 (1.44x) |
|    4096 |  128000 |       127000 |         2720.67 | 1935.24 (1.41x) | 1207.38 (2.25x) |

[Ubospica]

LGTM. Thanks for the fix!