[REQUEST] Allow eval batch size in pipeline

When I use PipelineModule.eval_batch(eval_iter) in my code, I found it necessary to set eval dataloader's batch_size same as train_micro_batch_size_per_gpu. For example, I set train_micro_batch_size_per_gpu to 2, and set eval dataloader' batch_size to 8:

val_dataloader = RepeatingLoader(DataLoader(val_dataset, 8))
eval_iter = iter(val_dataloader)

Then run train_batch and eval_batch:

for i in range(epochs):
    model.train_batch()
    model.eval_batch(eval_iter)

Then at the end of the first stage, there is an output tensor e.g. hidden_state, whose shape is 2, 16, 1024. But at the second stage, its shape changed to 8, 16, 1024, which may cause some unexpected results.

reset_activation_shape can solve it, but in my opinion calling reset_activation_shape before and after evaluation is a bit weird...so I think it may be more convenient to support this feature. Sorry to bother :)

Start by tracing PipelineModule.eval_batch and the existing reset_activation_shape behavior, comparing them with train_batch and the eval_iter example in the issue. The change is complete when evaluation can use a different dataloader batch size without callers manually resetting activation shape, while outputs remain consistent across pipeline stages.