Try optimizing imageCleanTransparent

[appgurueu]

• Goal of the PR: Same as Different algorithm for imageCleanTransparent #14478
• How does the PR work? Simply implements a BFS, keeping the "border" of nodes in a vector. Resulting images should look the same (basic testing confirms this), except there is no arbitrary iteration cutoff (because this new algorithm runs in overall linear time, whereas in the previous algo each iteration was linear time, this should not be needed, but it could be added easily).
• Not benchmarked yet.

To do

This PR is WIP.

• See FIXMEs: Why is the existing code not gamma correct (should we preserve this for performance)? Why are we weighing by alpha, which "over-weighs" propagated colors?
• Benchmark
• Optimize this PR (or kill it if the benchmarks show that it's the same.

How to test

Hack to make all nodes opaque / normal drawtype (this lets you see what happens to apple / rose / etc. without needing RenderDoc and having to look for the textures):

minetest.register_on_mods_loaded(function()
	for name, def in pairs(minetest.registered_nodes) do
		if name ~= "air" then
			minetest.override_item(name, {
				use_texture_alpha = "opaque",
				drawtype = "normal",
			})
		end
	end
end)

[sfan5]

thought: so the Bitmap class is essentially a std::set<v2u32>. and you're using a the vectors like a set.
maybe it's faster to (re)use a bitmap here, because the bitmap is a fixed 32 bytes while the vector goes to 160 bytes if you store even 20 pairs in it?