Metal backend for pl_gpu — interest check + working branch

[simonchrz]

Hi — I've written a native Metal backend for pl_gpu and wanted to check
interest before filing a formal merge request.

Status: it's working, not a sketch. It mirrors the Vulkan/GL/D3D11 backends
(a pl_gpu_fns vtable + public creation API + CAMetalLayer swapchain), reusing
the in-tree GLSL emitter via GLSL → SPIR-V (pl_spirv) → MSL (SPIRV-Cross) →
MTLLibrary/PSO. ~2260 LOC over 13 staged commits.

What works:

• Passes the full gpu_tests suite (buffer/texture/shader incl. film grain, HDR
  peak, all transfer functions, dithering, error diffusion). 25 formats.
• plplay renders real H.264 through pl_renderer on Metal.
• pl_timer returns real GPU times (MTLCommandBuffer timestamps).
• Validated on a physical device (Apple A19 / iPhone 17 Pro), including the
  iOS-only packed format bgr565.

Branch (public, reviewable via the compare view):
master...simonchrz:libplacebo:metal-pr-series

Three questions before I invest more polish:

1. Do you want a Metal backend in-tree at all, or would you rather it lived
   out-of-tree? I'm happy to maintain the Apple side either way.
2. CI: the GitLab pipeline has no macOS runner, so the existing CI can't build
   this. How would you like to handle that — a macOS job (I can draft one), or
   merge-with-the-understanding-it's-community-tested?
3. Anything in the structure you'd want reshaped before a serious review.

I'll file the formal MR on code.videolan.org once my GitLab account is approved
(currently awaiting admin approval); opening here first since GitHub PRs are
restricted to collaborators. Happy to move the discussion to GitLab if you prefer.

[simonchrz]

Update: since GitHub PRs on this mirror are restricted to collaborators, I've opened the formal merge request on the canonical GitLab repo:

videolan/libplacebo!858 — https://code.videolan.org/videolan/libplacebo/-/merge_requests/858

Happy to keep the discussion wherever you prefer (here or on GitLab).

[simonchrz]

Status update: the GitLab MR (videolan/libplacebo!858) is rebased onto current master and now includes the production-hardening commits (private storage for GPU-only textures, single-command-buffer batching, external queue support). It has been the sole render path of a shipping iOS app for ~2 weeks (M1/M4/A16/A19, full gpu_tests green). Review welcome over there — happy to split the series into stages if that's easier.