[Bugfix] Fix shape calculation for group quantization

[kylesayrs]

Purpose

• Fix shape calculation for group quantization
• This was uncovered while attempting to quantize the vision tower of qwen2.5 model. This likely went unnoticed, since most model shapes are powers of two, something that is not true of weights in the qwen 2.5 vision tower

Changes

• Use the ceil of num columns / group_size in order to reflect that the remainder columns are a group

Testing

• The test added by Add warning for non-divisible group quantization vllm-project/llm-compressor#1401 fails without these changes, but succeeds with them

[dsikka]

Should verify effects on vLLM, if any as I think the same calculation is assumed there when creating parameters for weight loading

[kylesayrs]

@dsikka That seems to be the case
https://github.com/vllm-project/vllm/blob/1c2bc7ead019cdf5b04b2f1d07b00982352f85ef/vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py#L104-L108

I can only assume that this is a mistake on vLLM side too that hasn't been triggered, in the same way that it hasn't been an issue for LC

[rahul-tuli]

Approved pending we raise a warning on llmcompressor side, for incompatible shapes with vllm. Thanks for the diff