[patch] minor fixes for 0.0.2

[francoishernandez]

• additions in convert_HF from Inference server, lots of related changes #42 broke some back-compatibility (e.g. llama2), which should be restored here;
• two issues in the lora merging script should be fixed: proper conservation of inference settings + use the compute_dtype from the lora config (e.g. if we finetune with a different dtype than the original model, like fusedadam+fp16 for an originally bf16 model)

[francoishernandez]

Merging this one to facilitate some work in #45 without stemming future conflicts. Might push some additional patches later.