[BUG] `mx.cummax` / `mx.cummin` mishandle `NaN`: device-dependent (CPU ≠ GPU), leak ±∞, and disagree with NumPy

[JasonHonKL]

Describe the bug
The Metal scan kernels for cumulative maximum / minimum
(mlx/backend/metal/kernels/scan.h, structs CumMax / CumMin) implement the
combine step with a plain ordered comparison that never propagates NaN:

To Reproduce

Include code snippet

import mlx.core as mx
import numpy as np

x = mx.array([1.0, 3.0, float("nan"), 5.0, 4.0])

print("np.maximum.accumulate:", np.maximum.accumulate(np.array(x)))
print("mx.cummax [gpu]      :", np.array(mx.cummax(x, stream=mx.gpu)))
print("mx.cummax [cpu]      :", np.array(mx.cummax(x, stream=mx.cpu)))

Expected behavior

np.maximum.accumulate: [ 1.  3. nan nan nan]
mx.cummax [gpu]      : [1. 3. 3. 5. 5.]      # NaN silently dropped
mx.cummax [cpu]      : [ 1.  3. nan  5.  5.]  # NaN kept at its slot, not propagated