vllm: improve container support

[russellb]

This is a follow-up issue for #181.

The vllm inline inference adapter works in both the conda and docker stack types, but some features fail in the docker case because the base image does not include all necessary dependencies (some cuda libraries, in particular). The specific case that failed for me was trying to use tensor_parallel_size greater than 1. NCCL fails to initialize (nccl library isn't present).

I started working on this, but didn't get it working completely yet.

my WIP was here: russellb@3a61246