Rewrite mergekit-extract-lora

[cg123]

Now with better embedding handling, multi-gpu execution, and lazy loading/saving of tensors.

When extracting a LoRA from an 8B model, execution time goes from ~6 minutes down to 40 seconds with --cuda --multi-gpu on an 8-GPU machine.

Additionally, the --sv-epsilon flag can be used to set a tolerance for singular values to opportunistically reduce rank when the fine tuned difference is inherently lower rank.

Also reimplement a couple of merge methods using the @easy_define decorator and add some missing tests.