Fix autocast v fp16 + add benchmark config instructions in benchmark.md

[eolecvk]

• Fixed conditional logic when using autocast in benchmark code.
• Updated diffusers==0.3.0 to diffusers==0.4.2
• @chuanli11 : Any reason to not include torch and torchvision in the requirements?

Removing autocast does seem to reduce the latency and memory in a big way!
Here is what the new benchmark results for autocast v fp16 look like.

device	precision	autocast	runtime	n_samples	latency	memory
NVIDIA RTX A6000	single	False	pytorch	1	7.37	7.73
NVIDIA RTX A6000	half	True	pytorch	1	4.39	4.52
NVIDIA RTX A6000	half	False	pytorch	1	3.34	3.38
NVIDIA RTX A6000	single	False	pytorch	2	14.0	9.57
NVIDIA RTX A6000	half	True	pytorch	2	8.29	8.74
NVIDIA RTX A6000	half	False	pytorch	2	6.17	4.92
NVIDIA RTX A6000	single	False	pytorch	4	27.25	18.4
NVIDIA RTX A6000	half	True	pytorch	4	15.73	11.4
NVIDIA RTX A6000	half	False	pytorch	4	11.59	8.82

Also added benchmark config instructions in benchmark.md

[chuanli11]

@eolecvk seems torch and torchvision are included in the setup.py.

I don't know if @justinpinkney has particular reason for doing it that way.

[eolecvk]

@eolecvk seems torch and torchvision are included in the setup.py.

I don't know if @justinpinkney has particular reason for doing it that way.

It's because the original repo was setup that way:
https://github.com/CompVis/stable-diffusion/blob/69ae4b35e0a0f6ee1af8bb9a5d0016ccb27e36dc/setup.py

...but even in their case, they install pytorch and torchvision in conda's environment.yaml:
https://github.com/CompVis/stable-diffusion/blob/69ae4b35e0a0f6ee1af8bb9a5d0016ccb27e36dc/environment.yaml