Preserve batches when a distributed worker gets only one - #3097

What does this PR do?

Keep each sampled item intact when DistributedSamplerWrapper assigns zero or one item to a worker. For example:

from speechbrain.dataio.sampler import DistributedSamplerWrapper

sampler = DistributedSamplerWrapper(
    [[0, 1], [2, 3]], num_replicas=2, rank=0, shuffle=False
)
print(list(sampler))  # before: [0, 1]; after: [[0, 1]]

itemgetter returns the item itself for one index, so a single batch becomes individual indices and the data loader raises TypeError: 'int' object is not iterable. A scalar sampler also raises, and an empty assignment calls itemgetter without arguments. Yielding one item per assigned index handles all three cases and preserves the existing multi-item behavior.

Added a runnable doctest and eight regression cases for scalar/batch samplers, padding, empty inputs and drop_last. Six cases fail before the change. Sampler, dataloader and batching tests plus sampler doctests: 24 passed, one CUDA-only test skipped. pre-commit run -a passes.

Also exercised Brain.fit with a DynamicBatchSampler in two actual CPU/Gloo processes, using the shipped CTC example's data pipeline and eight speech recordings (18.96 seconds). With one four-utterance batch per worker, the baseline fails before the first update. The fix completes four updates per worker; losses, all 10 model-state tensors and all 27 optimizer tensors exactly match a control using explicitly assigned batches. This uses a recurrent-only GRU encoder to run DDP on CPU; it does not claim GPU or convergence validation.

No new dependencies or API changes.

PR review