[BUG] Sparse MaskedCategorical entropy includes padding or returns NaN

[AHMETHAKANBEZIR1]

Bug description

On current main 8f92a50cb1e8543d6d7545605e4e5b399224928b, sparse MaskedCategorical.entropy() does not exclude padded indices. With default neg_inf=-inf, a padded logit produces 0 * -inf = NaN. With finite neg_inf, padding incorrectly contributes to entropy instead of being excluded, unlike the dense-mask path and the method's documented valid-outcome behavior.

Reproduction

import torch
from torchrl.modules.distributions import MaskedCategorical

logits = torch.tensor([1., 2., 3., 4.])
indices = torch.tensor([0, 2, -1])
for neg_inf in (float('-inf'), -10.):
    dist = MaskedCategorical(logits=logits, indices=indices,
                             padding_value=-1, neg_inf=neg_inf)
    print(dist.entropy())  # NaN or an incorrect finite value
print(torch.distributions.Categorical(logits=logits[[0, 2]]).entropy())
# tensor(0.3653): entropy over the two valid actions

Environment: Windows, Python 3.12.14 / PyTorch 2.10 CPU, TorchRL Python source checkout, TensorDict 0.14.2 official wheel. The optional TorchRL C++ extension is not built; this reproduces in the Python distribution implementation.

I intend to mask padding/non-finite zero-probability contributions before the existing entropy normalization, and test entropy and gradients against independent compact torch.distributions.Categorical references. This is separate from sample/log_prob shape handling in #4511 and mode ties in #4509. Investigated with Codex assistance; reproduction run locally.

[AHMETHAKANBEZIR1]

/assign