feat(verl): add MigrationBench Megatron LoRA recipe

[luyuzhe111]

Summary

• add a MigrationBench GRPO recipe for Qwen3-Coder using Megatron, LoRA, AgentCore rollouts, and W&B
• add a Python 3.12 Megatron dependency group with NVIDIA Megatron-Bridge and CUDA 13 Transformer Engine wheels
• recognize Qwen3-Coder's chat template as the qwen3_5 tool-call schema

Validation

• ran the recipe end to end against the deployed migration agent runtime
• pytest -q tests/backends/experimental/verl (53 passed)
• pytest -q tests/rollout_gateway/test_response_schemas.py (7 passed)
• live Qwen3-Coder template-hash test (1 passed)
• uv lock --check and locked Megatron sync dry run
• pre-commit hooks

[Linbo-Liu]

Is this already in /examples/strands_migration_agent?

[Linbo-Liu]

https://github.com/awslabs/agentcore-rl-toolkit/blob/main/examples/strands_migration_agent/eval_utils.py

[luyuzhe111]

@Linbo-Liu ah good catch. some functionalities do overlap. the code here however also addresses verl dataset format compatibility requirements.

we should in-theory reuse these functions but they are not really importable from the trainer environment. but to avoid inconsistencies i added a docstring to this file. does this work?

[Linbo-Liu]

SG!