[Feature]: Custom attention masks

[ojus1]

Inspired from this paper, we're exploring ways to bootstrap a bidirectional-context LLM from a decoder-only Causal LLM (e.g. llama-3). This is very easy to do in huggingface transformers by passing a custom attention mask.

Looking for guidance on how to make this happen in vLLM? TLDR;

1. Compute bidirectional hidden states from prompt.
2. Use causal attention for decoding. Help appreciated!