[WIP] Using a Compute graph as alternative to Observables

[SimonDanisch]

This introcudes ComputePipeline (all names very WIP) as an alternative to Observables, which solves:

• some performance problems
• multi input/output situations
• updating multiple attributes at once
• thread safety and better clean up
• Hopefully better stacktraces with clearer input/output mappings
• Maybe better compile times and inference

State of this PR: first implementation for Scatter, hacking into the plot pipeline to replace the Scatter constructor and hijack it with a computegraph implementation for convert_arguments/convert_attributes and any backend computation.
Only prototyped for GLMakie right now.
It currently leaves the plot type unchanged and stores the compute graph in plot.args[1] (lol) and overloads a few methods already just for Scatter (not getproperty etc though, yet).

Next steps:

• axis dim converts
• get/setproperty
• more benchmarks
• WGLMakie/CairoMakie test
• Same treatment for lineplot, which should really help see any performance gains

Continuation of: #4550 and #4360

[ffreyer]

If I understand correctly this interprets nothing as "the value has not changed". In that case shouldn't it leave the dirty state untouched? If a previous call has dirtied output i and this call has not changed it, it should still dirty, right?

Suggested change

	if isnothing(value)
	edge.outputs_dirty[i] = false
	else
	edge.outputs_dirty[i] = true
	edge.outputs[i][] = value
	end
	if !isnothing(value)
	edge.outputs_dirty[i] = true
	edge.outputs[i][] = value
	end

[SimonDanisch]

In that case shouldn't it leave the dirty state untouched? If a previous call has dirtied output i and this call has not changed it, it should still dirty, right?

No, the dirty state is local to the current resolve.
So needs to be set to the correct value when setting the new value.
We still need to double check if there's a simpler and more consistent way and add some tests if this works 100% as intended, but right now outputs_dirty is only for the children of an edge to figure out after a resolve if a value has changed from that particular resolve.

[ffreyer]

Progress on GLMakie:

The number of failing tests is now down to 15, with 27 tests throwing errors.

TODOs/Problems:

• Legend & Colorbar integration
• dim_converts integration
• SpecAPI integration
• dynamic px_per_unit changes (e.g. through save) (this may not work with camera matrix caching as it is implemented here)
• scatter may not match billboarding behavior of master, as rotation isa Billboard never happens. Needs tests on master to compare
• axis3d!() sometimes draws wrong, maybe a transformation issue?
• transform overwrite don't work yet

[ffreyer]

CairoMakie is now at 2 fails - 28 errors, GLMakie at 14-27