add parallel sampling using vllm

[daspartho]

close #370

adds support for parallel sampling using vllm library when num_return_sequences in generation kwargs is > 1 and the model is supported by vllm (currently all hf models in llm-vm)

TODO: handle dependencies

[VictorOdede]

Let's set vllm_support to be a property of the BaseOnsiteLLM class then we can set it to true by default unless it's not supported by the model then we can set the property to false

[daspartho]

made suggested changes. vllm_support is set to true by default and needs to be set false explicitly for unsupported models.