perf(h3): overlap the 2-GPU recipe rollout with engine concurrency 2

[Rockdu]

What

• --sglang-server-concurrency 1 → 2 in the 2-GPU MiniMax H3 t2va recipe.

Why

At concurrency 1 the engine has nothing queued while a finished sample is encoded and
handed back, so it idles between denoise runs. 2 lets the next sample's denoise overlap
that handoff.

Validation

Measured on this recipe, 2x H200, identical rollout_batch_size x n_samples_per_prompt
(2 x 8 = 16 videos per rollout), medians over 3 rollouts:

concurrency	run	rollout phase	train_wait	step_time	throughput
1	diffusion_grpo_h3_t2va_260820-205715-990	730 s	785 s	1263 s	0.022 samples/s
2	diffusion_grpo_h3_t2va_260821-021535-704	445 s	502 s	986 s	0.036 samples/s

• pre-commit run --files scripts/run_diffusion_grpo_h3_t2va_2gpu.py passes.
• python3 scripts/run_diffusion_grpo_h3_t2va_2gpu.py --help still parses.

Files

File	Role
scripts/run_diffusion_grpo_h3_t2va_2gpu.py	recipe: the flag change plus its rationale

Checklist

• pre-commit run --all-files passes — run on the touched file only
• Added/updated tests for new behaviour — no tests added; the change is a launch flag in an example script
• pytest -x is green — not run
• If launch flags changed, python3 train.py --help still parses — verified on the recipe entrypoint
• If a public flag was added, it appears in the CLI reference docs — n/a, no new flag, only a changed value
• If an example was added, it has a real walkthrough — n/a, existing example