sd: sync with master-758-c674225

[wbruna]

master-749-b11c95a:

• layer and row splitting between devices

[wbruna]

I've changed the internal C++ API to keep it more aligned with upstream, moving the backend, params and max_vram selection logic to the Python level; so we may enable the multi-device support without additional API logic.

[wbruna]

master-758-c674225:

• a fix for layer splitting with unet models
• use larger tiles for VAE tiled encoding
• DPM++ 2M SDE samplers
• denoise strength as starting noise level (optional)
• stream model conversion (only relevant for sdmain)

The denoise strength tweak would be a welcome addition for image to image mode, since it avoids the step count lowering when reducing the strength; except it deviates from the auto1111 api behavior 😕 Although I'm not sure if the 'send as refimg' toggle isn't breaking compatibility already (if it is, an approach could be making it a tri-state strength-as-step-count / strength-as-noise-level / strength-ignored-for-ref-image).