Guard the MLX convolution's rank at compile time

dispatch/mlx/pooling.py guards the rank it can handle and raises at compile time with a message naming the backend. dispatch/mlx/conv.py has no such guard, so a rank-4 convolution compiles cleanly and fails inside MLX on the first call, with a message that names neither the layer nor the backend policy. Every other NotImplementedError in the dispatch tree fires during pytensor.function(); this is the only one that waits until data loading, initialization and mx.compile are all behind you.

import numpy as np
import pytensor
import pytensor.tensor as pt

from pytensor_ml.layers.conv import ConvLayer

X = pt.tensor("X", shape=(1, 4, 4, 4, 4, 2))
W = pt.tensor("W", shape=(2,) * 6)
f = pytensor.function([X, W], ConvLayer((2,) * 4, (1,) * 4, (1,) * 4)(X, W), mode="MLX")
# compiles fine

f(np.zeros((1, 4, 4, 4, 4, 2)), np.zeros((2,) * 6))
# ValueError: [conv] Only works for inputs with 1-3 spatial dimensions.

One rank check mirroring the one already in the sibling pooling module turns this into a compile-time error like every other backend hole.

Compare the rank guard in dispatch/mlx/pooling.py with dispatch/mlx/conv.py, then use the issue's pytensor.function() reproduction as the first check. Done means unsupported convolution ranks fail during compilation with a message naming the layer and backend policy, rather than on the first call.