chore(verl): upgrade to verl 0.9.0 and vLLM 0.24.0

[luyuzhe111]

Summary

• Replace the git-pinned verl snapshot with the verl==0.9.0 release and upgrade vLLM from 0.23.0 to 0.24.0, which includes the fix from vLLM #44483.
• Keep transformers==5.12.1 for the rollout gateway parser while scoping the dependency override to verl instead of overriding the environment globally.
• Simplify the example launch configs by removing defaults and legacy flags, and select the Triton MoE backend required by the migration-agent setup.

Validation

• MigrationBench Megatron LoRA E2E completed 30 training steps with no rollout failures, aborted responses, OOMs, NCCL errors, or tracebacks. Checkpoints and validation completed successfully at steps 15 and 30.
• Validation reward remained comparable to the previous stack:

Step	Previous	Upgraded	Delta
0	0.4130	0.4264	+1.34pp
15	0.5786	0.5619	-1.67pp
30	0.6756	0.6522	-2.34pp

Both runs show the same improving trend, with the upgraded stack within 2.4 percentage points through step 30.