Add deepseek support

[hello-cpaxton]

Description

Adds deepseek as a model option under qwen. Example usage:

python -m stretch.app.chat --llm qwen25-Math-7B-Int4
python -m stretch.app.chat --llm qwen25-0.5B-AWQ
python -m stretch.app.chat --llm qwen25-Coder-7B
python -m stretch.app.chat --llm qwen25-Deepseek-1.5B

Something such as

python -m stretch.app.chat --llm qwen25-Deepseek-3B

should raise error as this is a non existing weight configuration

Note that Deepseek fine tunes aren't all avaialble (i.e., there doesn't seem to be a 3b). So I added bitsandbytes support for quantization, and you can now run Qwen 7b or 14b on ~5 or ~10 gb of vram:

python -m pip install bitsandbytes qwen_vl_utils autoawq transformers>=4.49.0

Checklist

• I have performed a self-review of my code
• If it is a core feature, I have added thorough tests
• I have added documentation for the changes
• I have updated the README file if necessary
• I have run on hardware if necessary

Screenshots (if applicable)

Add any relevant screenshots or screen recordings to help reviewers understand the changes.

Additional context

Add any other context or information about the pull request here.

[hello-peiqi]

Based on my search, DeepkSeek R1 only distills Qwen of size 32B, 14B, 7B, and 1.5B

[hello-peiqi]

• Coder: 32B, 14B, 7B, 3B, 1.5B, 0.5B, (None, Instruct, Instruct-AWQ, Instruct-GGUF, Instruct-GPTQ-Int4, Instruct-GPTQ-Int8)
• Math: 72B, 7B, 1.5B (None, Instruct)
• VL: 72B, 7B, 3B (Instruct, Instruct-AWQ)
• None: "0.5B", "1.5B", "3B", "7B", "14B", "32B", "72B" (None, Instruct, Instruct-AWQ, Instruct-GGUF, Instruct-GPTQ-Int4, Instruct-GPTQ-Int8)
• Deepseek: 1.5B, 7B, 14B, 32B