Fix extract block for tensor spaces

[jorgensd]

To be able to easily extract blocks from ufl.MixedFuntionSpaces.

Motivation:
Instead of writing out every block of a and L by hand, as

    a00 = ufl.inner(ufl.grad(u), ufl.grad(v)) * dx
    a01 = ufl.inner(c, v) * dx
    a10 = ufl.inner(u, d) * dx

    L0 = ufl.inner(f, v) * dx + ufl.inner(g, v) * ufl.ds
    L1 = ufl.inner(zero, d) * dx

    a = dolfinx.fem.form([[a00, a01], [a10, None]], dtype=dtype)
    L = dolfinx.fem.form([L0, L1], dtype=dtype)

one can now do

    dx = ufl.Measure("dx", domain=mesh)
    f = -ufl.div(ufl.grad(u_ex))
    n = ufl.FacetNormal(mesh)
    g = ufl.dot(ufl.grad(u_ex), n)
    a = ufl.inner(ufl.grad(u), ufl.grad(v)) * dx
    a += ufl.inner(c, v) * dx
    a += ufl.inner(u, d) * dx
    L = ufl.inner(f, v) * dx + ufl.inner(g, v) * ufl.ds
    L += ufl.inner(zero, d) * dx

    a_blocked = ufl.extract_blocks(a)
    L_blocked = ufl.extract_blocks(L)

    a = dolfinx.fem.form(a_blocked, dtype=dtype)
    L = dolfinx.fem.form(L_blocked, dtype=dtype)

It also fixes extract_blocks for tensors, and properly document this function, as to what happens if i or j is not provided.
Now it extracts the blocks as a scalar, vector or tensor.

[mscroggs]

This change shouldn't be in this PR (see #307)

Suggested change

	for i, j in enumerate(np.ndindex(e.reference_value_shape)):
	for i, j in enumerate(np.ndindex(e.value_shape)):

[jorgensd]

I'll revert it once 307 is merged.

[cdaversin]

Looks good to me :)

[jpdean]

Nice! This'll be really handy for coupled problems. Do we have a test?

[jorgensd]

Nice! This'll be really handy for coupled problems. Do we have a test?

I think we can adapt the stokes demo.

I use this in many of my own works lately.