Bump vllm and transformers dependencies

[lewtun]

Hello again, the pinned versions of vllm and transformers are quite outdated and as a result it is not simple to run newer models / archs with PipelineRL. If I recall correctly, one reason to pin vllm was due to the generator / trainer mismatch, but this can largely be solved with either (https://blog.vllm.ai/2025/11/10/bitwise-consistent-train-inference.html), upcasting the LM head to FP32 or applying importance sampling.

It would be nice if the deps could be pinned to the latest compatible versions with the codebase :)

[rafapi]

Hi @lewtun, Thanks for highlighting this. We're currently in the process of upgrading our dependency stack as this issue is currently affecting our work internally.