Expose DeviceDescriptor::memory_hints via WGPUDeviceExtras

[RubyBit]

Checklist

• cargo clippy reports no issues
• cargo doc reports no issues (unrelated to my changes)
• cargo deny issues have been fixed or added to deny.toml
• Human-readable change descriptions added to CHANGELOG.md under the "Unreleased" heading.
  • If the change does not affect the user (or is a process change), preface the change with "Internal:"
  • Add credit to yourself for each change: Added new functionality. @githubname

Description

wgpu's DeviceDescriptor has had memory_hints since MemoryHints was introduced, but map_device_descriptor hardcoded Default::default() (Performance) to be changed in the future (so I changed it): Performance hints make gpu-allocator commit 128 MiB device + 64 MiB host memory blocks for the few tiny internal allocations wgpu makes at device open, so creating a bare device commits ~200 MiB of VRAM (measured on NVIDIA RTX 4080 Laptop/Vulkan; D3D12 is similar), and all committed memory rounds up in 256 MiB steps. Applications with small or mostly-static workloads have had no way to opt out through the C API (such as my tensor library which can load pretty small models on the gpu through wgpu and there is an unnecessary, around 200 MiB overhead).

Add a WGPUMemoryHints enum (Undefined/Performance/MemoryUsage/Manual) and memoryHints + suballocatedDeviceMemoryBlockSize{Start,End} fields to WGPUDeviceExtras, mapped onto wgt::MemoryHints. Zero-initialized extras keep the previous behavior (Undefined -> Performance).

Measured bare-device committed VRAM (RTX 4080 Laptop, Vulkan):
Performance (old default): 201 MiB
MemoryUsage: 21 MiB
Manual { 4 MiB..256 MiB }: 17 MiB

Related Issues

[inner-daemons]

Added by #594 so unless you need this instantly I would prefer to wait.

[RubyBit]

I see, but when can we expect the PR merged then? v30 also released a few weeks ago.

[inner-daemons]

@RubyBit I would expect within the next month at most for it to be merged, but I have hopes it could happen sooner.