MLX AdvancedIncSubtensor fails when Blockwise adds a batch dim

The MLX dispatch for AdvancedIncSubtensor does x.at[indices].add(y) with the unbatched indices, so a scatter-add that vectorize_graph has given a leading batch dim can't broadcast. Works on every other backend.

import numpy as np
import pytensor
import pytensor.tensor as pt
from pytensor.graph.replace import vectorize_graph

B, N, K = 4, 7, 3
idx = pt.constant(np.random.default_rng(0).integers(0, K, size=N))
levels = pt.vector("levels", shape=(K,))
grad = pt.grad(levels[idx].sum(), levels)  # AdvancedIncSubtensor: scatter-add into (K,)

batched = pt.matrix("batched", shape=(B, K))
vectorized = vectorize_graph(grad, replace={levels: batched})
x = np.ones((B, K), dtype=pytensor.config.floatX)

print(pytensor.function([batched], vectorized, mode="FAST_RUN")(x).shape)  # (4, 3)
pytensor.function([batched], vectorized, mode="MLX")(x)
# ValueError: [broadcast_shapes] Shapes (4,1) and (7,3) cannot be broadcast

Start with the MLX dispatch for AdvancedIncSubtensor and reproduce the vectorize_graph example from the issue in MLX mode. The fix is complete when the batched scatter-add runs without a broadcast error and returns an array with shape (4, 3), matching the other backend.