[Feature] Update apply mask kernels

[Ubospica]

Previously the apply mask kernel computes:

for i in range(len(indices)):
    for j in range(vocab_size):
        if get_bitmask_value(bitmask, i, j) == 0:
            logits[indices[i], j] = -inf

This means the shape of the bitmask is (len(indices), bitmask_size), where len(indices) represents the number of structured generation requests in a batch. However, this number can be different for different batches, and the shape of the bitmask also differs.

This PR makes the shape of the bitmask consist across batches to make it easier to allocate the bitmask. It fixes the shape of the bitmask to (batch_size, bitmask_size), and modifies the apply mask kernel to:

for batch_id in indices:
    for j in range(vocab_size):
        if get_bitmask_value(bitmask, batch_id, j) == 0:
            logits[batch_id, j] = -inf

GPU Kernel benchmarks:

_______________________________ test_apply_token_bitmask_inplace_large[True-1-128000-1024-1] ________________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 5.824641790241003 us
______________________________ test_apply_token_bitmask_inplace_large[True-1-128000-120000-1] _______________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 6.118565797805786 us
______________________________ test_apply_token_bitmask_inplace_large[True-1-128001-120000-1] _______________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 5.8456724509596825 us
______________________________ test_apply_token_bitmask_inplace_large[True-1-128010-120000-1] _______________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 5.760767031461 us
_______________________________ test_apply_token_bitmask_inplace_large[True-64-128000-1024-1] _______________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 19.754506647586823 us
______________________________ test_apply_token_bitmask_inplace_large[True-64-128000-120000-1] ______________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 62.413290143013 us
_______________________________ test_apply_token_bitmask_inplace_large[True-64-128000-1024-4] _______________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 20.551515743136406 us
______________________________ test_apply_token_bitmask_inplace_large[True-64-128000-120000-4] ______________________________
--------------------------------------------------- Captured stdout call ----------------------------------------------------
apply_token_bitmask_inplace_cuda time: 33.12954679131508 us

It also removes CUDA kernel impl.