speculative decoding support

[natehofmann]

build off of liangjason87:liangjason/spec_decoding, but with rough Rust layer. haven't tried yet, things still all over the place. keeping to simple, non-stream, non-lora case

[mmoskal]

I think the spec-decoding logic should sit in async_exec.rs and not be visible outside much.

• disable logits_processor on the draft model for now
• spin two threads for response from trtllm (you already do that)
• in ReqData keep the current draft and real req id

When a request from user is added (assume n_draft_tokens=5)

• enqueue request in draft executor, with max_tokens=5
• when the draft responder thread gets the response (all 5 tokens), validate tokens against the constraint (may ignore for now), and enqueue new request in the real executor (max_tokens=1, though maybe that's implicit)
• when the real responder thread gets the response (I guess it's going to be up to 6 tokens), send StepResults back to the user and enqueue new request in draft executor (if request not done)

This way you don't have to mess with completions.rs much.