Faster copy for col contig to row contig

[awni]

Adds a specialized copy from col-contiguous to row-contiguous arrays.

Benchmark on an H100:

Pre: 818 GB/s
Post: 2191.62 GB/s

The benchmark is below:

import time
import mlx.core as mx

t = mx.bfloat16
D = 4096
x = mx.random.normal(shape=(D, D)).astype(t)

def fun(x):
    for _ in range(20):
        x = mx.contiguous(x.T)
    mx.eval(x)

for _ in range(20):
    fun(x)

tic = time.time()
for _ in range(20):
    fun(x)
toc = time.time()

gb = 20 * 20 * x.nbytes * 2 / 1e9
s = toc - tic
gbps = gb / s
print(f"{gbps=:.3f}")

[angeloskath]

Looks great! Left a couple of comments.