Batch skinned meshes

[superdump]

Objective

• Make it possible to batch/instance skinned meshes

Solution

• Skinned meshes were excluded from batching/instancing because each skinned mesh entity has its own joint matrix 'skin uniform' binding of a uniform buffer at a dynamic offset. As such, each entity has to be drawn separately in order to set the dynamic offset for each draw.
• This PR adds support for using a storage buffer for the joint matrices where available.

---

Changelog

• Enable batching/instancing of skinned meshes

[JMS55]

@nicopap @superdump we should figure out how to combine this with #9902

[JMS55]

Suggested change

	pub fn from(mesh_transforms: &MeshTransforms, skin_index: u32) -> Self {
	pub fn new(mesh_transforms: &MeshTransforms, skin_index: u32) -> Self {

[nerdachse]

Agree with this change, as from has a special meaning in Rust wrt the From-trait (I am aware that you're aware, but potential new contributors might not)

[JMS55]

Suggested change

	let is_storage = render_device.limits().max_storage_buffers_per_shader_stage > 0;
	let is_storage = render_device.limits().max_storage_buffers_per_shader_stage > 0;

At this point, I think we should add a method like "storage_buffers_supported()" to RenderDevice.

[superdump]

What does your suggestion change? The lines look identical to me.

[JMS55]

I didn't change anything sorry. This was meant to be a regular comment on that line.

[nicopap]

Have you considered defining this struct as a standalone type in bevy_render? Similarly to GpuArrayBuffer? Even if it is only used here, it helps separating concerns. It will also make it easier to remediate conflicts with #9002 (independently of who has to fix the merge conflicts)

[superdump]

The problem is that it's very specific to Mat4 at the moment. Ideally it feels like this should be a different 'mode' of GpuArrayBuffer. One mode for batching as many instances of data into one binding, and one where you have a known max binding size that you use for a uniform buffer, but you may not use it all so you align to the next dynamic offset binding as needed and make sure the last dynamic offset has a full binding of data after it. That requires more thought to implement.

[superdump]

And if it weren't specific to Mat4, we could use an affine matrix instead and save space.

[superdump]

I figured out how to do this, and would like to. But I think it has to wait until after 0.12. I'll move this to 0.13.