Relates to huggingface#1525

Unfortunately, the docs I wrote about how to use LoftQ were incorrect,
based on a misunderstanding I had. In reality, it is quite a bit more
involved to get LoftQ working, requiring a complete roundtrip first
loading a non-quantized model with LoftQ, saving the LoRA weights and
the modified base model, loading the just stored base model again but
this time with quantization, and finally loading the LoftQ-initialized
adapter on top. The docs now link to the example which demosthenes how
to move through these steps.

Add reco to use all-linear and nf4.