Support block_bucketize_sparse_features for empty tensor

[JacoCheung]

Bug description

While evaluating with TorchRec ShardedEmbeddingCollection under a distributed env, some rank may get zero input for the last global batch. However fbgemm.block_bucketize_sparse_features will throw errors when the lengths tensor contains no elements:

Traceback (most recent call last):
    x = torch.ops.fbgemm.block_bucketize_sparse_features(
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/_ops.py", line 1120, in __call__
    return self._op(*args, **(kwargs or {}))
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: invalid configuration argument
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Can you help to support an empty input ? Thanks!

reproducer

import fbgemm_gpu
import torch
import torchrec

values = torch.tensor([]).long().cuda()
lengths = torch.tensor([]).long().cuda()

mb = torchrec.KeyedJaggedTensor(
    keys = ["product"],
    values = values, # key [] on rank0, [2] on rank 1
    lengths = lengths, # length [] on rank0, 1 on rank 1
)
x = torch.ops.fbgemm.block_bucketize_sparse_features(
  lengths,
  values,
  bucketize_pos=False,
  sequence=True,
  block_sizes=torch.tensor([2]).cuda(),
  my_size=2,
  max_B=-1
)

ENV

• version 1.0.0