Fix metal scan

[awni]

Fix race condition in metal kernel. It's quite difficult to get this to show up in a unit test.

Here's a test that triggers it every so often.

import mlx.core as mx

mx.random.seed(0)
int_arr = mx.random.randint(0, 4, (4000000,))
arr = mx.array(int_arr)
for i in range(200):
    print(mx.cumsum(arr))

And in a separate file:

import subprocess

command_arr = []
for i in range(100):
    command_ = "python test_mlx_cumsum.py"
    command_arr.append(command_)

# start all programs
processes = [subprocess.Popen(command_, shell=True) for command_ in command_arr]
# wait
for process in processes:
    process.wait()

[angeloskath]

Great catch thanks!