docs: Clarify multi-gpu usage

[tgaddair]

trafficstars

Using --gpus all for docker run also requires --sharded or --gpus N to be set for LoRAX, but this isn't made clear. We should add something in the docs about GPUs and using multi-GPU.

[tgaddair]

Also, should add some docs explaining tensor parallelism, and when it makes sense to use multi-GPU. Specifically at least one of:

• Model is too big for one GPU
• GPUs are connected via NVLink

Otherwise the network overhead of GPU-to-GPU communication will be the main bottleneck.