[Tests] Small Fixes

[dsikka]

Summary

• Update flakey test. Periodically produces NaNs - maybe from recently updating the dtype to bfloat16?
• Don't cast zp back to dense dtype post rounding

[kylesayrs]

Thanks! I think this is a lot more correct and prevents any unintended effects on zero points.

We should consider also not casting to original dtype during quantize, but we can think about that more later.

https://github.com/vllm-project/compressed-tensors/blob/main/src/compressed_tensors/quantization/lifecycle/forward.py#L477-L479