fix issue of long tensor in torch gather in interpretable input (#1678)

Summary: Pull Request resolved: #1678 `torch.gather` requires `index` as `long` in the released versions of pytorch `int32` is already supported in source code (why the internal tests pass) this diff gonna fix [the CI fails](https://github.com/meta-pytorch/captum/actions/runs/19831857827/job/56826961913), which is caused by test cases using int32 `mask` Reviewed By: jimshao1999 Differential Revision: D88101395 fbshipit-source-id: fc37972c0437eff03abc5250e53a59ad7fc02317

Author: aobo-y

captum/attr/_utils/interpretable_input.py

@@ -55,6 +55,11 @@ def _scatter_itp_attr_by_mask(
  else:
  expanded_itp_attr = itp_attr
 
+ # gather index must be long
+ # may support int32 soon https://github.com/pytorch/pytorch/pull/151822
+ if expanded_feature_indices.dtype != torch.long:
+ expanded_feature_indices = expanded_feature_indices.long()
+
  # gather from (*output_dims, *inp.shape[1:-1], n_itp_features)
  attr = torch.gather(expanded_itp_attr, -1, expanded_feature_indices)