MLX backend: every pad raises TypeError (pad_width passed as mx.array)

[guillaume-osmo]

Description

mlx_funcify_pad forwards pad_width straight to mx.pad, but the linker typifies every input to mx.array while mx.pad takes an int or a list of (before, after) int pairs, so no pad works on this backend. Same root cause as #2386.

test_mlx_pad[constant_default] and test_mlx_pad[edge] are already red on main.

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

x = pt.matrix("x", shape=(3, 4), dtype="float32")
xv = np.zeros((3, 4), dtype="float32")

print(pytensor.function([x], pt.pad(x, pad_width=2, mode="edge"), mode="CVM")(xv).shape)  # (7, 8)
print(pytensor.function([x], pt.pad(x, pad_width=2, mode="edge"), mode="MLX")(xv).shape)
# TypeError: pad(): incompatible function arguments
#   pad_width: int | tuple[int] | tuple[int, int] | list[tuple[int, int]]

Note the value cannot simply be read back at runtime: the linker enables mx.compile by default and MLX forbids evaluating a traced array, so a constant pad_width has to be resolved at funcify time, as in #2386.