Unsloth fixes Mistral Medium 3.5

This is the kind of release note that matters more than a flashy benchmark bump: if you run models locally, compatibility bugs can quietly tank results even when the weights are fine.

• –The root cause was a YaRN parsing quirk, with `mscale_all_dim` changing from `1` to `0` resolving the bad behavior.
• –The fix touched multiple runtimes, so users on transformers or llama.cpp should refresh their files instead of debugging their own pipelines first.
• –Updated GGUFs and corrected `mmproj` output make this more than a narrow bug patch; it improves the practical deployability of the model.
• –It also shows how model quality now depends on the surrounding inference stack, not just the checkpoint itself.