Enable Ragged Rollouts: Allow Variable-Length list[sample] per Batch

Could we support padding-style samples—i.e., samples that are masked out so they don’t participate in weight updates or advantage computation? The main use case is when each rollout returns a variable number of samples. My current approach is to pad with dummy samples to force every rollout to return a fixed-length list[sample].

Even better would be supporting variable-length list[sample] within a batch. For example, with rollout_batch_size = 1 and n_samples_per_prompt = 2, the first rollout might return one sample and the second rollout two samples—giving 3 in total, which can’t be evenly divided by 1 * 2

This happens with longer conversations, where one sample has to be split into multiple samples. Consequently, the length of the list[sample] returned by a single rollout becomes variable.

The issue does not name files, tests, or entry points. Start by tracing how rollouts are batched and how list[sample] lengths are validated, then define how padding masks should affect weight updates and advantage computation; done means variable-length samples work within a batch without dummy samples or uneven-size failures.