Replace torch image by vllm image in Docker

[qgallouedec]

Note that if we want to use vllm in the images for e.g. GRPO, a better approach is to use the official vllm image like this:

FROM vllm/vllm-openai:v0.10.2

Originally posted by @lewtun in #4522

[lewtun]

FYI this makes the image significantly larger (about 10GB vs ~3GB), but has the advantage of packaging the right torch version with the vllm one