[feat] Add prompt_mean loss reduction

[erictang000]

Add a prompt_mean option to algorithm.loss_reduction: compute the token-mean within each prompt group (the n_samples_per_prompt responses sampled for a prompt), then average over prompts. Each token [i, t] in prompt p is scaled by 1 / (num_prompts * tokens_in_prompt_p) so that summing the per-token policy loss yields mean_p(token_mean within prompt p). Unlike token_mean, every prompt contributes equally regardless of total token count.

• preprocess: add compute_prompt_boundaries(uids) for per-prompt slices (works for step-wise and non-step-wise training).
• trainer: thread per-prompt boundaries through metadata into _normalize_advantages, rebased to mini-batch-relative indices.
• ppo_utils: implement the prompt_mean branch in apply_loss_reduction_to_advantages_minibatch.
• config validation + docs updated for the new option.
• unit tests for compute_prompt_boundaries and prompt_mean.

[gemini-code-assist]

Code Review

This pull request introduces a new loss reduction strategy called prompt_mean, which computes the average token loss within each prompt group and then averages over all prompts. The changes include implementing the prompt boundary computation, updating the advantage normalization logic in the trainer, adding configuration validation, and providing comprehensive unit tests. There are no review comments, so I have no feedback to provide.