2d: Use dynamic buffers for instance data

[stuartcarnie]

Important

Depends on

• Add Persistent Buffers and Push Constants #103779

Summary

This PR updates the 2D canvas renderer to utilise dynamic buffers for instance data, per the changes in 7861cdd.

I modified the Canvas 2D renderer to use dynamic buffers for the InstanceData (similar to what @darksylinc did for forward / mobile renderers). This change removes two copies, so the instance data is written once (on UMA devices):

• one to upload the InstanceData to the GPU, and
• a second to BLIT the instance data from the staging to the private GPU storage buffers.

I modified the 1x1 sprite benchmark in @Calinou's comment, to draw 20,000 textured rectangles, which still batches, but reduces the CPU bottleneck of rendering 20,000 individual canvas items, to demonstrate the advantages of this PR. TileMaps is an example that will benefit from these changes.

Modified code

extends Node2D

const ITERATIONS = 20_000

@onready var spr: Sprite2D = $Sprite2D as Sprite2D

var draw_instances := false

func _ready() -> void:
	# Use fixed seed for reproducibility.
	seed(0x60d07)

	if draw_instances:
		for i in ITERATIONS:
			var sprite: Sprite2D
			sprite = spr.duplicate() as Sprite2D
			sprite.position.x = randf_range(0, 1152)
			sprite.position.y = randf_range(0, 648)
			add_child(sprite)

func _draw() -> void:
	if !draw_instances:
		var tex: Texture2D = spr.texture
		var pos: Vector2 = Vector2.ZERO
		for i in ITERATIONS:
			pos.x = randf_range(0, 1152)
			pos.y = randf_range(0, 648)
			draw_texture(tex, pos)

About a 10% improvement. The average FPS went from 3161 FPS to 3487 FPS

[stuartcarnie]

Merged into #111183