fix(perf): Don't use data-dependent op when batch_size and max_num_nodes are provided to to_dense_batch

[akihironitta]

Summary

This PR addresses two issues in to_dense_batch:

1. Currently, to_dense_batch triggers a D2H sync even when batch_size and max_num_nodes are provided because of num_nodes.max() > max_num_nodes to decide whether to run boolean masking on x and idx.
2. In addition, the boolean masking is a data-dependent op which causes a graph break. An alternative would be to enable capture_dynamic_output_shape_ops=True, but in our case, this new implementation guarantees that the shape is static for a given batch_size and max_num_nodes at compile time, so there's no need to continue to use the op.

Benchmark

• Env: torch==2.12.0.dev20260322+cu130, g6.4xlarge (L4)
• Code: https://gist.github.com/akihironitta/836b95989930f8e9312deef65d89d3a8

Before (3a4b881)

[---------- bs=1024 mn=64 ----------]
                     |  cpu   |  cuda
8 threads: --------------------------
      eager          |  44.2  |  4.3
      overflow       |   9.2  |  2.9
      torch.compile  |  41.9  |  3.7

Times are in milliseconds (ms).

[---------- bs=1024 mn=512 ----------]
                     |   cpu   |  cuda
8 threads: ---------------------------
      eager          |  465.1  |  24.4
      overflow       |  252.9  |  14.7
      torch.compile  |  455.5  |  23.8

Times are in milliseconds (ms).

[---------- bs=1024 mn=2048 ----------]
                     |   cpu    |  cuda
8 threads: ----------------------------
      eager          |  1670.7  |  95.9
      overflow       |   941.0  |  56.2
      torch.compile  |  1651.8  |  93.2

Times are in milliseconds (ms).

After (this PR)

[---------- bs=1024 mn=64 ----------]
                     |  cpu   |  cuda
8 threads: --------------------------
      eager          |  34.7  |  2.8
      overflow       |  16.3  |  2.5
      torch.compile  |  33.3  |  1.4

Times are in milliseconds (ms).

[---------- bs=1024 mn=512 ----------]
                     |   cpu   |  cuda
8 threads: ---------------------------
      eager          |  290.3  |  16.8
      overflow       |  214.0  |  12.4
      torch.compile  |  294.4  |  15.5

Times are in milliseconds (ms).

[---------- bs=1024 mn=2048 ----------]
                     |   cpu    |  cuda
8 threads: ----------------------------
      eager          |  1096.5  |  67.1
      overflow       |   827.8  |  49.6
      torch.compile  |  1062.9  |  63.0

Times are in milliseconds (ms).