perf: reduce map picker memory with a page-backed thumbnail cache

[robertoesteves13]

Summary

Adds a page-backed (unmanaged) thumbnail cache for map picker backgrounds and removes the two things that made the picker grow the process by hundreds of megabytes. Measured on Windows x64, 2880x1800 @175%, the real app window, the real picker and a 95-mapset / 87-background library.

Commits

• feat: add OS page allocator for unmanaged bitmap buffers — new Utils/PageAllocator.cs (VirtualAlloc/mmap), AllowUnsafeBlocks in the csproj (required by LibraryImport).
• feat: add page-backed thumbnail cache with Avalonia decode and resize — new Utils/BitmapStorageHandler.cs: Load(path) decodes with Bitmap.DecodeToWidth, center-crops to the 2:1 page and blits straight into a pre-allocated page; path-keyed cache hits skip decode and blit entirely.
• feat: serve map picker card backgrounds from the thumbnail cache — one line in SongMapsetCardViewModel.BuildBackgroundImage.
• perf: virtualize the map picker mapset list — ItemsPanel StackPanel → VirtualizingStackPanel.
• perf: freeze the modal backdrop blur instead of animating a live effect — ModalHost no longer keeps a live BlurEffect on the window shell.

Why

1. A live BlurEffect over WindowShell (which contains the whole app UI), animated 0→10 px over 260 ms, with a BitmapCache. Every composited frame re-blurs a full window-sized surface (~20 MB at 2880x1760) and the native allocator never returns it. Because the effect stays attached while the modal is open, any animation above it re-blurs the window — that also covers the picker's compositor-driven smooth scrolling.

2. The picker list was not virtualized. ItemsControl + StackPanel realized every loaded card (~1.2 MB each) and paging appends 20 more per scroll, so memory grew with the library.

Measurements

phase	before	after
first modal open, peak delta	+208 MB	+70…79 MB
2nd / 3rd open	—	+3.4 MB / +1.0 MB
scroll, 20 → 80 cards loaded	+53.4 MB, unbounded	+16 MB, realized containers pinned at 6–9
picker open / close (app window, main)	—	438 MB → 517 MB → 421 MB (released on close)
87 backgrounds through the cache, standalone	—	+18 MB peak, decode transients freed

For context, an empty maximized window in a bare Avalonia app on this display already measures 245.9 MB (77 MB before the window) — that floor, not the picker, is most of the process footprint.

Trade-offs / behaviour changes

• The backdrop blur is no longer animated: it is applied in one step at open (the backdrop fade and the card slide/fade are unchanged) and is frozen while a modal is open — content behind it does not animate. It is re-taken if the window is resized while open.
• The thumbnail cache is keyed by path only, so if a background file changes in place while its page is cached, the stale thumbnail is served. Easy to extend with an mtime check in BitmapStorageInfo.
• Pages are process-wide (BitmapStorageHandler.Shared), committed lazily on first use (20 × 512 KiB) and returned to the OS at exit.
• With VirtualizingStackPanel the scroll extent is estimated while further pages load; worth an eyeball on a very large library.
• Headers in the tool views and BeatmapSourceCard still use ArtworkBitmapUtils.DecodePreview at 800/1200/480 — unchanged here.

Notes for review

• Avalonia has no zero-copy path from caller-owned memory to a renderable bitmap: new Bitmap(PixelFormat, AlphaFormat, IntPtr, …) copies into an SKBitmap, and the uncopied route (protected Bitmap(IBitmapImpl)) needs IDrawableBitmapImpl, which is internal to Avalonia.Skia. So a Load costs one 512 KiB memcpy; the cache saves the decode + crop, not the copy.
• The frozen backdrop is captured with two 1:1 RenderTargetBitmap.Render(Visual) passes (blur applied through an off-tree Image with BlurEffect). Scaling a render target through its DPI is not reliable on this window: with the extended-client-area window the content is drawn 1:1 into the smaller bitmap, which shows up as a magnified, cropped backdrop.
• #if windows was never defined (SDK defines WINDOWS, and only for OS-specific TFMs), so the UNIX allocator branch was the one every Windows build ran. Also fixed: MEM_LARGE_PAGES can never be satisfied by a 512 KiB page (needs 2 MiB granularity + SeLockMemoryPrivilege), and VirtualFree must be called with MEM_RELEASE alone and dwSize == 0.

Testing

No permanent tests added — MapWizard.Tests has no Avalonia host. Verified with throwaway harnesses (deleted) that drove the real MainWindow, the real picker view model and the real song library:

• page store byte-identical to the returned bitmap; cover-crop correct in both branches (2:1 source and panorama); JPEG and PNG; cache hit after the source file was deleted; 23 images over a 20-page round-robin; allocator alloc/free/zeroing/edge writes.
• picker card lifecycle through SetBackgroundActive; 25-card scroll sweep; reactivation served from the cached page rather than a re-decode (file replaced in place to prove it).
• paired screenshots of the live app (picker open vs closed) confirming the blurred backdrop is pixel-aligned with the live UI and that the shell restores cleanly on close.

dotnet build is clean (one pre-existing MainWindow.axaml AVLN3001 warning).

Untested here: the UNIX mmap path (no Linux/macOS available).

[maotovisk]

Unfortunately needed to revert the blur optimization: it was VERY clever but it didn't respect window geometry (was looking very bad on wayland) + animations, so i added some toggles for animations and motion on the appearance tab, will revisit some form of it in the future tho.

Also to make you happy i added a GruvBox palette.