[BUG] QubitUnitary decomposition fails with qjit for the 2 CNOT case

[dime10]

Originally reported in the forum (https://discuss.pennylane.ai/t/error-when-using-catalyst-grad-with-blockencode/9234), the following code will fail in the latest release:

import pennylane as qml
import catalyst
from jax import numpy as jnp

A = jnp.array([[1, 0], [0, 1]])

dev = qml.device("lightning.qubit", wires=2)
@qml.qnode(dev)
def loss_fn(angle):
    qml.BlockEncode(A, wires=[0, 1])
    qml.RZ(angle, wires=0)
    return qml.expval(qml.Z(0))

angle = jnp.array(0.5)

grad_fn = catalyst.qjit(catalyst.grad(loss_fn))
grads = grad_fn(angle)

The numpy.ndarray conversion method __array__() was called on traced array with shape float64[4]

The issue is improper detection of the abstract array case for U with the current Catalyst frontend, located here:

pennylane/pennylane/ops/op_math/decompositions/unitary_decompositions.py

Lines 216 to 219 in dce345c

	# The 2-CNOT decomposition relies on sorting eigenvalues, which is not supported
	# with abstract tracers when capture is enabled. In that case, we fall back
	# to the 3-CNOT decomposition.
	if not capture.enabled():

I believe the bug was introduced in PR #8717