form splitter: use DAGTraverser and handle Action, Adjoint

[leo-collins]

Makes the split_form MultiFunction be a DAGTraverser.

Also handles splitting Adjoints and Actions. Previously, something like

W = V * V
u, _ = TrialFunctions(W)
I = interpolate(u, V)
c = Cofunction(V.dual())
F = action(adjoint(I), c)
split_form(F)

would fail because F has a single argument numbered 0, but the inner Interpolate has an argument numbered 1, so ExtractSubBlock would try to select a non-existant block.

For Adjoint, we swap the row and columns, for Action we need to maintain the argument that gets contracted out.

[pbrubeck]

This is great. Should we consider migrating this to ufl?

[pbrubeck]

This is great. Should we consider migrating this to ufl?

Splitting a Function might be the only real blocker because we do things with MixedDat.

[pbrubeck]

Suggested change

	return Argument.reconstruct(a, function_space=subspace(a.function_space(), indices))
	return a.reconstruct(function_space=subspace(a.function_space(), indices))

[leo-collins]

This called the Firedrake Argument reconstruct which complains about the value shape not matching

[pbrubeck]

I would be tempted to remove that check

[pbrubeck]

Why can we not map(self) anymore?

[pbrubeck]

DAGTraverser.__call__ is doing something else?

[leo-collins]

blocks is no longer an attribute on self, so we have to pass it down. Before, _subspace_argument() was called in the split() method when handling a ZeroBaseForm.

[pbrubeck]

Maybe this works

from itertools import repeat

map(self, o.arguments(), repeat(blocks))

[leo-collins]

Calling self on the arguments would return a split argument. ZeroBaseForm needs unsplit arguments on the subspaces.

[pbrubeck]

I resolved a merge conflict after #4093 got merged