[utsav1033]

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: I used AI to find the bug with a CPU/GPU comparison sweep, it built MLX and ran tests and gave me the test code.

Proposed changes

Sorting a bool array on the GPU fails, while the same call works on the CPU:

import mlx.core as mx

a = mx.array([True, False, True, False])
mx.sort(a, stream=mx.cpu)  # works
mx.sort(a, stream=mx.gpu)  # fails: Unable to load kernel c_block_sort_bool__bool__bn32_tn4

The cause is that sort.metal instantiates the sort kernels for every dtype except bool, so no bool kernel is ever built. This PR adds bool_ to both instantiation lists:

• the single-block sort, used for small rows
• the multi-block sort, used for large rows (sorted in chunks, then merged)

i didnt change anything else as sorting relies on comparison and its worknig for bool(false<true), and on a padding value, which in turn exists as Limits::max = true.

argsort, topk and partition use the same kernels, so this fixes them for bool too.

Testing

• Added test_sort_bool to test_ops.py. It fails before this change and passes after.
• Checked against the CPU and numpy for sizes from 3 to 1M, which covers both the single-block and multi-block paths. The checks covered the argsort indices, topk, partition, 2D and sliced inputs, and all-True, all-False and empty arrays.
• All 177 tests in test_ops.py pass, and the formatting checks pass.

Tested on my Apple M4, on current main