Extend AsBindGroup macro with support for RenderAssetUsages

[davids91]

What problem does this solve or what need does it fill?

When having large amount of data used in a compute shader, which does not change frequently; It does not need to be uploaded to the GPU in every loop.

e.g.:

#[derive(Resource, Clone, AsBindGroup, TypePath, ExtractResource)]
#[type_path = "MyAmazingData"]
pub struct MyAmazingData {
    #[storage_texture(0, image_format = Rgba8Unorm, access = ReadWrite)]
    pub output_texture: Handle<Image>,

    #[uniform(1, visibility(compute))]
    pub frequently_changing_data: Vec3,

    #[storage(2,  visibility(compute), read_only)]
    pub(crate) large_data: Vec<u32>,
}

What solution would you like?

Extend the AsBindGroup macro with optional RenderAssetUsage attributes: render_usage(...) where the options render_world or cpu_world where the default value would be cpu_world.

#[derive(Resource, Clone, AsBindGroup, TypePath, ExtractResource)]
#[type_path = "MyAmazingData"]
pub struct MyAmazingData {
    #[storage_texture(0, image_format = Rgba8Unorm, access = ReadWrite)]
    pub output_texture: Handle<Image>,

    #[uniform(1, visibility(compute))]
    pub frequently_changing_data: Vec3,

    #[storage(2,  visibility(compute), read_only, render_usage(render_world))]
    pub(crate) large_data: Vec<u32>,
}

What alternative(s) have you considered?

I have looked everywhere to make this happen, and have not found any way to do this; The closest to a solution for this
was extending the concept described in #8440

Additional context

Here is a GPU profile for a voxel raytracer using bevy:

The actual rendering does not take up a significant amount of time, instead the voxel data is being copied in every frame to the GPU.

The function taking ~500ms in each loop is only calling as_bind_group, here