GEM: Simplify Indexed tensors

[pbrubeck]

Introduces simplifications that destroy indexed structures

ListTensor([Indexed(A, i) for i in numpy.ndindex(A.shape)]) -> A

Indexed(ListTensor(array), FixedIndex) -> array[FixedIndex]

As a benchmark, we generate code for assembling

from firedrake import *
from firedrake.tsfc_interface import compile_form
V = FunctionSpace(UnitTetrahedronMesh(), "JM", 1)
v = TestFunction(V)
u = TrialFunction(V)
form = inner(u, v) * dx + inner(div(u), div(v)) * dx(degree=0)
kernels = compile_form(form, "stokes", parameters={"mode": "spectral"})

With this PR it takes 9.41 seconds. Using current main, this takes around 20 minutes, the majority of which is spent in sum_factorise.

[ksagiyam]

I marked this PR as Draft as you still seem to be debugging. Please mark it as Ready for review when you become confident.

[pbrubeck]

Tests are passing (modulo known doc/linkcheck failures) https://github.com/firedrakeproject/firedrake/actions/runs/17322729298/job/49186947155

@ksagiyam I feel quite happy with this, could you review?

[ksagiyam]

I'm not sure if this is safe. This is a recursion, and, unlike when we use DAGTraverser, the result is not cached.

[pbrubeck]

What is the issue? Are we potentially creating new objects and not freeing memory for the old ones?

[pbrubeck]

I moved this case to the remove_componenttensors DAG traverser.

[connorjward]

I think that this needs some more comments explaining what is going on. To someone not quite up to speed with the intricacies of GEM this is very hard to understand.

Otherwise I think I'm quite happy with this.