torch.compile and profile

[mahmoudhas]

Refactors the pipeline to use standalone functions for different steps and applies torch.compile where possible. Observed the following improvements using torch.compile:

• the UNet forward pass went down from 200 ms to 100 ms (per denoising step)
• the VAE forward pass went down from 250 ms to 110 ms (per inference chunk)
• The ImageProcessor (face and mask detection) time went down from 290 ms to 230 ms (per inference chunk)

The compilation takes around 1 minute.

TODO for production readiness:

• re-use the python process to serve multiple requests, saving compilation time.
• remove the torch profiler
• pad the last inference chunk so it doesn't trigger a recompilation