AsyncGRPOTrainer: add model_init_kwargs support

[rycerzes]

What does this PR do?

Adds model_init_kwargs support to AsyncGRPOTrainer, matching the existing behavior in GRPOTrainer.

Currently, AsyncGRPOTrainer hardcodes AutoModelForCausalLM.from_pretrained(model, device_map=None, dtype=torch.float32), which prevents passing custom kwargs like attn_implementation or dtype. This PR:

• Adds model_init_kwargs field to AsyncGRPOConfig (with _VALID_DICT_FIELDS for dict serialization support)
• Replaces direct AutoModelForCausalLM loading with create_model_from_path(model, **model_init_kwargs), which infers the correct architecture class from the model config (important for models like Qwen3.5 that use Qwen3_5ForConditionalGeneration instead of a standard CausalLM)
• Enforces device_map=None for FSDP2 compatibility

Changes:

• async_grpo_config.py: add model_init_kwargs field + docstring + _VALID_DICT_FIELDS
• async_grpo_trainer.py: replace AutoModelForCausalLM with create_model_from_path, update imports

Tests:

• test_model_init_kwargs: verifies that model_init_kwargs={"dtype": "bfloat16"} results in bfloat16 model parameters

Closes part of #5831

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

AI writing disclosure

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Who can review?

@qgallouedec

---

Note

Low Risk
Small, localized change to model instantiation with an existing shared helper and a focused unit test; no auth or rollout pipeline logic changes.

Overview
AsyncGRPOTrainer can now take optional model_init_kwargs on AsyncGRPOConfig, aligned with GRPOTrainer, so string model IDs accept from_pretrained options (e.g. dtype, attn_implementation) instead of a fixed AutoModelForCausalLM load with dtype=torch.float32.

Loading goes through create_model_from_path, which picks the architecture from the hub config (not only causal LM). device_map=None is always applied for FSDP2. A test asserts bfloat16 parameters when model_init_kwargs={"dtype": "bfloat16"}.

^{Reviewed by Cursor Bugbot for commit b925002. Bugbot is set up for automated code reviews on this repo. Configure here.}