Add wgpu-mojo

[Hundo1018]

Add wgpu-mojo — pure Mojo bindings for wgpu-native (WebGPU).

Compute and graphics on every backend Vulkan / Metal / DX12 supports, from a single
Mojo program: RAII wrappers that release their native handle on scope exit,
strongly-typed handle newtypes instead of raw pointers, a high-level GPU facade
that does compile + dispatch + read-back in a few lines, and a GLFW-backed
RenderCanvas for windowed rendering.

• Source: https://github.com/Hundo1018/wgpu-mojo
• Release: https://github.com/Hundo1018/wgpu-mojo/releases/tag/v0.2.1
• License: Apache-2.0
• Maintainer: @Hundo1018

Checklist

• My recipe.yaml file specifies which version(s) of MAX is compatible with my project. — This package does not depend on max at all. It pins mojo-compiler =1.0.0 in build/host and uses pin_compatible('mojo-compiler') at run time, per the packaging guide. The MAX interop bridge lives in a separate top-level package that is deliberately excluded from the build, precisely so max stays off every consumer's dependency path.
• License file is packaged. — license_file: LICENSE; verified present as info/licenses/LICENSE in the built artifact.
• Set the build number to 0. — New package.
• Bumped the build number (if the version is unchanged). — N/A, first submission.

Notes for reviewers

Platforms. The recipe skips everything except linux-64 and osx-arm64,
including the linux-aarch64 this repo's CI builds. Nothing in the binding is
architecture-specific and conda-forge ships wgpu-native for aarch64, but no one
has run wgpu-mojo there, so I would rather not publish a package nobody has
exercised. I checked that the skip expression resolves correctly on all five
targets, so the aarch64 leg exits cleanly rather than failing your build.

Native dependency. wgpu-native and glfw are declared host/run
dependencies resolved from conda-forge — the build script downloads nothing. It
compiles the two C callback bridges from source in the repo (wgpu-native's async
API needs real C function pointers, which Mojo cannot yet produce) and links
against the host environment.

ABI pin. wgpu-native is pinned exactly, not as a range. Upstream renumbered
the whole 0x0003xxxx SType enum in the v29.0.0.0 → v29.0.1.1 patch release; a
mismatched library still loads and still runs, it just misreads every extras
chain. Beyond the version constraint, conda.recipe/build.sh compares the
installed $PREFIX/include/wgpu.h byte-for-byte against the copy vendored in the
source repo and fails the build on any difference.

CodeQL is enabled on the source repository (it contains C and Python
alongside Mojo) and the badge is in the README.

Test. test_package.mojo runs in a bare environment and asserts three things:
that lib/mojo/wgpu.mojoc is discoverable without an -I flag, that both native
libraries load, and that the installed wgpu-native exports the drift-prone entry
points this binding targets. It requests no adapter, so it passes on a build
machine with no GPU and no display server.

Verification performed

Built from this exact recipe and revision, then installed from a local channel
into a throwaway pixi project with no source checkout on the path, no -I flag
and no inherited LD_LIBRARY_PATH:

==> pixi add wgpu-mojo
✔ Added wgpu-mojo >=0.2.1,<0.3

==> Assertion 1: package resolves, native libraries load
wgpu preflight OK
  wgpu-native version: 486539264
  symbol check: OK (25 critical symbols present)
  adapters found: 4

==> Assertion 2: compute round trip on a real adapter
compute round trip OK — all 1024 elements correct

The same was repeated against the published .conda downloaded back from the
GitHub Release, and the package test passes standalone via
rattler-build test --package-file. Both checks run in the source repo's CI on
ubuntu-latest and macos-14.