Run NR before DLSS upscaling, with per-pass controls and an RR option

[wilsjo2]

This lets NR run before DLSS upscaling instead of after it. For example, at 4K with DLSS Performance, NR works on the 1080p image and DLSS upscales the result to 4K. That reduces the work NR has to do, with some loss of fine detail compared with running it at full resolution.

It also adds:

• One, two or three NR passes, with separate history for each pass.
• Style, intensity, structure, tone and skin controls for each pass. Later passes can inherit the first pass's settings.
• A separate option for NR after native Ray Reconstruction. RR keeps its original inputs; NR runs on its output. The pass count and model resolution for this path are separate from normal DLSS.
• A fallback to post-upscale NR when the input uses padded or offset render areas that the pre-upscale path cannot safely handle.

The preset selectors are still available under an advanced section. I haven't confirmed that they visibly change the model, so style is the main selector.

The pre-upscale two-pass path was tested in BG3's DX11-to-DX12 mode: 1080p NR followed by DLSS to 4K. The latest per-pass controls and RR path compile, but still need more in-game testing. I haven't verified this with RTX 40 MFG unlockers.

I saw #6, #23 and #26 cover some of the same ground. This is the version I've been using in my fork, offered here in case the combined controls and fallback handling are useful. Happy to split out the parts you want.

This also includes my fork's README, install notes and packaging changes. Those aren't needed for the rendering changes and can be left out if you prefer. No NVIDIA model DLL is included.