Add BindGroupLayout caching via descriptors

[Zeophlite]

Objective

• Defer creating BindGroupLayout by using a BindGroupLayoutDescriptor and cache the results
• Unblocks bevy_material (render-less material definitions)

Solution

• Reviewers, look at first commit for mechanism, and following for usage

Testing

• CI

[Zeophlite]

This is yuck, AsBindGroup::unprepared_bind_group needs a BindGroupLayout , so we fetch it from the cache - investigate if unprepared_bind_group can use BindGroupLayoutDescriptor instead

[Zeophlite]

TODO

[atlv24]

I think you can do similar to what the method right above this one does, something like

Suggested change

	todo!();
	BindGroupLayoutDescriptor::new(
	Self::label(),
	&Self::bind_group_layout_entries(render_device, false),
	)

[atlv24]

Ohh, the entries method takes render_device. Hmm. im not sure why that is

[Zeophlite]

Seems to be related to #[derive(AsBindGroup)] macro with bindless , it needs the render_device to see if its supported

[Zeophlite]

See Zeophlite@1c9e74c - this disables the above check, but hopefully helps point in the right direction

[Zeophlite]

Okay I think the bindless_supported check can be deferred but too much for this PR, I've added back render_device for bindless cases, see 3bae3f4

[atlv24]

@Zeophlite I think the way we should go about this is to remove the render device dependency in the other direction: we don't want to defer it but make it happen earlier, and inform as bind group by some other mechanism. For example, if we had a way of setting shader defs like BINDLESS and then checking them in as bind group, we can generate the right bind group layout etc for it. Because thats really what it is, shader-def-dependent bind group layouts.

[atlv24]

I think for the time being using a render device to implement it is fine though :)

[alice-i-cecile]

Can you add more motivation to the PR description please? I'm having trouble following why this is a good idea.

[atlv24]

Reviewers: this line (and the equivalent on ComputePipelineDescriptor) is the whole point of this change. It makes these two structs not depend on wgpu, which means they can be moved out of bevy_render into a bevy_material, and created without the need of a renderdevice. Specifically, this allows describing a shader's inputs without bevy_render, which when combined with bevy_shader allows for renderless material abstraction.

Everything else in this PR is just to make these two line diffs happen.