Fix grammar logits processor

[SwadicalRag]

This pull request addresses:

1. The function signature of GrammarLogitsProcessor does not match with the current implementation of _apply_logits_processors in sampler.py
2. The use of Ray to filter logits results in a significant slowdown on my local machine. Given performance is the priority, and I haven't seen any drawbacks to doing so, I have removed the grammar module's dependency on Ray.

[AlpinDale]

Thanks for the PR! The RayRemote logits processor was added for multi-gpu scenarios where Ray is involved. Have you tested the PR with Ray enabled?

[SwadicalRag]

Yes I have. I run a 2x 3090 system, and the performance was halved or worse with ray. That being said, I was hammering the server exclusively with grammar constrained requests.

Without grammar on a 7b model, I get well over 100 Tok/s. With grammar and ray, only 20-30 Tok/s. With ray removed 60 Tok/s.

The overall bottleneck is the latency in waiting for lark to return valid tokens. I suspect there are latency issues with ray doing multiple copies of tensors/objects and synchronizing across threads too.