Proposal: clarify naming of rollout/training loop CLI arguments

[DavidBellamy]

Summary

The three core sampling-training loop arguments all use "rollout" but mean different things, making them hard to reason about (especially the constraint equation). Proposing clearer alternatives.

Current names vs. proposed

Current	Proposed	Rationale
--rollout-batch-size	--prompts-per-cycle	This is the number of prompts per sampling cycle, not a training batch size
--num-steps-per-rollout	--train-steps-per-cycle	Number of optimizer steps run on one cycle's data (default 1 for on-policy)
--num-rollout	--num-cycles	Total iterations of the sample-then-train loop
--n-samples-per-prompt	--responses-per-prompt	"Samples" is overloaded in ML; these are generated completions
--global-batch-size	--train-batch-size	Clarifies this is samples consumed per optimizer step

Why this matters

The constraint equation becomes self-documenting:

# Before
rollout-batch-size × n-samples-per-prompt = global-batch-size × num-steps-per-rollout

# After
prompts-per-cycle × responses-per-prompt = train-batch-size × train-steps-per-cycle

The "after" version reads as plain English: "total responses generated per cycle = total responses consumed per cycle."

Suggestion for backwards compatibility

The old names could be kept as deprecated aliases so existing scripts don't break.