Improve fused_assemble design

[charleskawczynski]

Right now, fused_assemble generates code like:

pairs = ()
  pairs = (pairs..., :(@. x += 1))
  for i in 1:2
    pairs = (pairs..., :(@. x += 2))
    pairs = (pairs..., :(@. x += 3))
  end
end

But this is type-unstable because the type of pairs changes across the loop. We could instead lower this to:

pairs = ()
  pairs = (pairs..., :(@. x += 1))
  pairs = (pairs..., Iterators.flatmap(1:2) do i
    (:(@. x += 2), :(@. x += 3))
  end
end

which the compiler may have a chance at inferring.

One consideration is: what if we have if-blocks inside? It should work with Union-splitting, but this approach might be problematic if there are runtime checks.

[charleskawczynski]

Also, if check_restrictions passes for simple internals, we can lower to a simple tuple(...) result.