DLSS-NR: dynamic chained model passes with per-pass overrides, plus an optional Chinese language pack

[Moeblack]

Runs the DLSS neural-rendering model more than once over the same frame on D3D12. Each extra pass is a separate NGX feature with its own frame history, fed the previous pass's answer, so later passes see one frame per frame instead of re-feeding one instance's own output. The pass count is not capped at a product number: 2-3 is the sensible range, the UI suggests 1-6, and a direct numeric input accepts any positive uint32_t. Off by default: [DlssNr] Passes=1 and IndividualPassSettings=false take the same path the code takes today (one evaluate with identical arguments; no scratch, no extra features, no chain).

This is an independent implementation on the same base as #23 (and alongside the draft in #1). The machinery is necessarily close family — separate per-pass features built a frame ahead so nothing is created and evaluated on one command list, a two-texture work set, modelInput left untouched so the resolve's edit = model - proxy stays the whole chain's. Where it differs: no fixed upper bound (extra features are created one per Dispatch, one at a time, and every build frame returns before evaluation); sparse per-layer overrides with live master inheritance; and the menu shows the count that actually completed (DlssNr::ActivePassCount()), not the requested one. If you would rather keep one multipass implementation in the tree I am happy to reconcile or rebase this onto whichever you pick.

Per-pass sparse configuration

[DlssNr] IndividualPassSettings=false by default: every layer uses the master model controls and nothing is copied or materialised. Setting it true does not snapshot the master values — each [DlssNr.PassN] section stores only the fields actually overridden; everything else keeps inheriting the master settings live. Sections are 1-based, may skip numbers, and lowering Passes never deletes higher overrides. On save the owned sections are rebuilt from the in-memory map so cleared fields and alternate spellings (Pass01) cannot resurrect stale keys. Override-map reads/writes are guarded by a dedicated mutex (the render side only looks up, never inserts).

Pipeline notes

• Extra features live in a dynamic vector; [0] is pass 2 by the renderer's counting (config uses 1-based pass numbers; the conversion is explicit at the single integration point).
• Work textures: the existing output plus outputPingPong, allocated only when Passes > 1; passes alternate UAV/SRV ownership as they read the previous answer and write the next. finalOutput is tracked by pointer, never re-derived from parity at resolve time.
• A pass whose settings changed is rebuilt; changing the upstream count or resolution resets the chain's history rather than recreating every downstream instance blindly.
• A failed extra pass stops the chain at the last good surface, latches a reason shown in the menu (with Retry); lowering the count clears the latch — no per-frame reallocation/retry storm.
• Native Vulkan keeps today's single-pass master behaviour; the menu says so instead of showing controls that would not apply. The DlssNrUseProxy path does not participate in the chain or per-pass overrides and is refused with an explanation.

Localization (self-contained; drop the last commit to remove entirely)

Optional, non-invasive Chinese language pack. English source text remains the key; Localization::Tr() looks up a UTF-8 dictionary (OptiScaler_zh.ini / zh_CN.ini next to the DLL, then beside the game exe) loaded once per DLL lifetime; missing files or keys fall back to the original English text, and translations that do not preserve every printf directive or that alter ImGui ##/### identity suffixes are rejected at load. When the pack is in use, the menu font atlas merges %WINDIR%\Fonts\msyh.ttc (resolved via GetWindowsDirectoryW) so CJK glyphs render instead of boxes. Without the pack nothing changes: no dictionary, no font merge, no behavioural delta.

All C++/H follows CONTRIBUTING.md PCH rules (new Localization.cpp includes pch.h first; no header includes it).

What I did not do

No HLSL/cbuffer or precompiled-shader changes; no NGX forwarder ABI change; no Vulkan multi-pass; no new model DLL; no new GPU queues. No VRAM formula is claimed beyond the host side (one extra work texture; each feature's internal state still grows with the count).

Test plan

• Release x64 MSBuild builds clean
• D3D12: Passes=1 picture is unchanged from today's build (default-identical path)
• D3D12: 1->2->3 live switching; even/odd final-output parity verified in GTA V Enhanced
• D3D12: per-pass overrides inherit master when IndividualPassSettings=false
• D3D12: menu shows the actually-completed layer count
• zh pack: with OptiScaler_zh.ini present the menu renders Chinese (incl. glyphs); without it the menu is byte-identical English