Apply lora by model patching

[StAlKeR7779]

Rewrite lora to be applied by model patching as it gives us benefits:

1. On model execution calculates result only on model weight, while with hooks we need to calculate on model and each lora
2. As lora now patched in model weights, there no need to store lora in vram

Results:
Speed:

loras count	hook	patch
0	~4.92 it/s	~4.92 it/s
1	~3.51 it/s	~4.89 it/s
2	~2.76 it/s	~4.92 it/s

VRAM:

loras count	hook	patch
0	~3.6 gb	~3.6 gb
1	~4.0 gb	~3.6 gb
2	~4.4 gb	~3.7 gb

As based on #3547 wait to merge.

[lstein]

I like it. Much more straightforward than the previous solution.