KV Cache Quantization support deepseek v3

[MichoChan]

[HDCharles]

This is a known issue on deepseek. I'll go through details and mark this a good first issue.

compare deekseekv2 and v3 to qwen

deepseek is using a different attention architecture due to MLA which our kv_cache attention quantization detection helper1 and helper2 doesn't detect.

theoretically its as easy as detecting these alternative names.

1. update compressed-tensors to detect these names. This shouldn't require changes to llmc or vllm as far as i can tell.
2. test end to end with deepseek v2/v3. You can edit the existing kv_cache examples and test there or here there was a PR which tried to add a test for this which would probably be good to add as well.