[BUG] `argsort` diverges between cpu and gpu on complexes with NaN

Describe the bug
An array with complex64 values containing NaNs sorts differently on gpu and cpu streams.

To Reproduce

import mlx.core as mx
NAN = float("nan")
a = mx.array([complex(1, NAN), 1 + 1j, 2 + 0j], mx.complex64)
mx.argsort(a, stream=mx.cpu) # -> [0 1 2]
mx.argsort(a, stream=mx.gpu) # -> [1 2 0]

GPU streams behave correctly (input is sorted, NaN last, result matches np.argsort)
CPU streams return [0,1,2,...n] without sorting.

Expected behavior
Expected outputs to match. Ideally they should behave as numpy et al.

Desktop (please complete the following information):

• OS Version: MacOS 27.0 RC
• Version : main (v0.32.2 + d9add9d11)

Additional context
Hit in PR #4501
Issue 100% written by me, a human.

Start by reproducing the complex64 NaN case with the Python mx.argsort entry point on both mx.cpu and mx.gpu streams. Trace the CPU and GPU argsort paths, then add or update coverage for the shown input so both outputs match numpy with NaN last; the issue is done when CPU and GPU results agree.