[Mistral][SWA] Add Mistral sliding window support

[CharlieFRuan]

This PR adds support for Mistral. The implementation follows the Mistral paper, specifically including sliding window attention (SWA), rolling buffer cache, and chunking, as discussed in Section 2 in the paper.

This PR is largely analogous to the changes in llm_chat.cc in mlc-llm's PR mlc-ai/mlc-llm#1087.

Different from the approaches in #202, this PR's implementation takes advantage of SWA, so that there is no max window size anymore--one of the main benefits of Mistral.

Tested:

• Works well with:
  • 4096 sliding window size, with 1024 chunk size
    • Note that the small chunk size has no effect on generated output, but only less memory requirement with slower speed
  • 2048 sliding window size with 2048 chunk size
• Tested with prompts that are multiple times in length of the sliding window size / chunk size

Irrelevantly, we also make wizard models reuse llama model libraries given the dynamic vocab size support (updated just now due to shuffle support).

cc @tqchen

[tqchen]

great, @CharlieFRuan let us make a new npm release and update the demo