feat: add a primitive form of continuous batching

[AlpinDale]

This PR adds a continuous batching path, for only modern GGUFs, and a select set of samplers. This is mostly a proof of concept, so a lot of features including context shift, smart context, and others will not work. To test:

python koboldcpp.py --model /path/to/model.gguf --multiuser 32 --continuous-batching 8 --noshift

--noshift is required, and --continuous-batching needs to be manually tuned; too high and it'll not work. Suffice to say, lots of work needs to be done. At least, it's really fast.

Mode	Requests	Concurrency	Max Length	Batch Slots	Total Wall Time	Throughput	p50 Latency	p95 Latency
Legacy queue	16	16	512	1	22.607s	0.708 req/s	11.970s	22.813s
Continuous batching	16	16	512	4	6.241s	2.564 req/s	3.824s	6.236s