[fix] Fix num_padding_microbatches metric for Megatron backend

[SumanthRH]

What does this PR do?

Fixes num_padding_microbatches metric calculation for the Megatron backend when there are real sequences with all zero loss masks (example with apply_overlong_filtering=true)

The number of padding microbatches should not count all sequences with zero loss mask because some of these are real sequences but masked out for loss calculation.

The fix is to read the num_padding_microbatches attribute from the microbatch_iterator when available.

[gemini-code-assist]

Code Review

This pull request updates the calculation of real and padding microbatches in megatron_worker.py. Instead of inferring the number of real microbatches by checking if the loss_mask sum is greater than zero, it now retrieves the padding count directly from the microbatch_iterator via num_padding_microbatches. This prevents real microbatches with an all-zero loss mask (e.g., due to DAPO overlong filtering) from being incorrectly classified as padding. Additionally, the status reporting is updated to use this padding count directly. There are no review comments, so no feedback is provided.