Automatically split fused kernels by parameter memory limits

[charleskawczynski]

This PR leverages CliMA/MultiBroadcastFusion.jl#52, so that we can have:

@fused_assemble begin # (or many broadcast expressions using `@fused_direct`)
    for i in 1:100000
        @. x += y * z
    end
end

And this will automatically fuse as many kernels as permitted by the available parameter memory.

[charleskawczynski]

Hm, let's split off the dependency updates