[transformer v2] Load + evaluate FID for OpenAI guided-diffusion models

[Birch-san]

• loads OpenAI guided-diffusion UNet
  • assigns to inner_model_ema only
  • constructs 256x256 guided diffusion UNet, with config.json overrides
  • prepares model_ema with accelerate
  • wraps it with k-diffusion wrapper
  • grabs sigma_min, sigma_max from wrapper
  • loads weights, requires_grad(False) + eval()s it
• --evaluate-only no longer requires you to specify a (redundant) --evaluate-every option
• --evaluate-only now turns off a bit more stuff
  • avoids constructing trainable model (so you don't pay RAM price of a copy you don't use, and so you don't pay VRAM to have accelerator prepare it)
  • avoids constructing + accelerator-preparing optimizer and schedules
  • doesn't force you to define a config for optimizer/schedules (well, there are defaults for those but still)
• switched demo/evaluate from no_grad to inference_mode, which is even lower-overhead

as for the config.json overrides for constructing the guided diffusion UNet…
compared to what was in your clip_guided_diffusion/main.py, I:

• changed image_size 512->256
• removed "rescale_timesteps": True, (making it default to False)
• removed "timestep_respacing": "1000", (making it default to "")
• removed "use_checkpoint": False, (which was already the default)

I did this to closer match the MODEL_FLAGS they documented in their README:
https://github.com/openai/guided-diffusion#classifier-guidance