[recipe, doc] feat: add MiniMax H3 FlowGRPO V1 sync recipes

[NancyFyong]

Summary

Add production MiniMax H3 FlowGRPO V1 sync launch recipes for both
supported GPU paths:

• run_minimax_h3_t2va_lora_v1.sh; and
• run_minimax_h3_fl2va_lora_v1.sh.

They are mechanical counterparts of the existing V0 recipes. Each preserves
all model, LoRA, reward, pipeline, and CPS FlowGRPO settings, and changes only:

main_diffusion -> main_diffusion_v1
trainer.use_v1=true
trainer.v1.trainer_mode=sync
experiment name *_v1_gpu

The MiniMax H3 README documents both launchers and explains that V1 sync uses
TransferQueue and ReplayBuffer. separate_async is intentionally out of scope:
it needs dedicated actor/rollout resource pools plus checkpoint-engine
synchronization, so it is not a mechanical entrypoint switch.

Why this is not duplicate work

The relevant V1 tracking issue is #389. Duplicate-work checks found V1 work for
Qwen-Image (#489) and OPD (#493 and #495), but no open PR or launcher for
MiniMax H3 FlowGRPO V1, T2VA, or FL2VA. In particular, no existing branch adds
run_minimax_h3_t2va_lora_v1.sh or run_minimax_h3_fl2va_lora_v1.sh.

This PR adds recipes only. It does not modify MiniMax H3 adapters, FlowGRPO
losses, the V1 trainer, or GPU smoke registration.

Tests

Passed:

bash -n examples/flowgrpo_trainer/minimax_h3/run_minimax_h3_t2va_lora_v1.sh
bash -n examples/flowgrpo_trainer/minimax_h3/run_minimax_h3_fl2va_lora_v1.sh

# Both scripts with --cfg job and representative environment overrides
# resolved trainer.use_v1=true and trainer.v1.trainer_mode=sync.

python -c 'import transfer_queue'
pre-commit run --files \
  examples/flowgrpo_trainer/minimax_h3/README.md \
  examples/flowgrpo_trainer/minimax_h3/run_minimax_h3_t2va_lora_v1.sh \
  examples/flowgrpo_trainer/minimax_h3/run_minimax_h3_fl2va_lora_v1.sh
git diff --check

A mechanical comparison also confirmed that each V1 launcher retains every V0
runtime setting after normalizing the approved entrypoint, V1 flags, and
experiment-name changes.

No GPU training run was performed: this PR is intentionally limited to
production recipes and does not add a GPU smoke. The existing TinyRandom
T2VA+FL2VA FlowGRPO E2E coverage is tracked separately in #447.

Reward Results

AI assistance and review

AI assistance (pi coding agent) was used for this change. The human submitter
has reviewed every changed line and is responsible for this contribution.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[knlnguyen1802]

For all v1 recipes, we need to verify the reward curve again

[NancyFyong]

For all v1 recipes, we need to verify the reward curve again

I will switch to "open" after the reward phase. Thankyou