Fix quantization_config parsing for --kv-cache-dtype=auto

[alvarobartt]

Description

Warning

This PR constitutes a breaking change given that as there was a bug before with the parsing order where dtype (and torch_dtype) where being parsed as the default --kv-cache-dtype when --kv-cache-dtype=auto (or simply not set); the actual dtype from the quantization_config was not being used as the default option, meaning that now models with a dtype or torch_dtype set with an invalid quantization_config will raise a RuntimeError whereas before those where working "fine", just estimating the KV cache requirements with a "wrong" dtype.

This PR fixes the placement of the quantization_config to have more priority if there than dtype or torch_dtype, as well as bumping the version to 0.4.4 with uv version 0.4.4.

---

• I have read and followed the guidelines in CONTRIBUTING.md.
• This has been discussed over an issue or discussion.