Explore vicuna-1.5 16k context models

[the-crypt-keeper]

Currently only transformers runtime can handle the fp16 versions, but vLLM has an open PR to support rope-scaling: vllm-project/vllm#555

Since we run everything else with vLLM it would be good to do apples-to-apples.

[the-crypt-keeper]

Given we now have good 32k and 100k models readily available, these older ones aren't worth the effort.