[Performance] mlx.core.conv_general is really slow

Describe the bug
Method mlx.core.conv_general is significantly slower than PyTorch analog. Can vary from 10x to 150x slower.

To Reproduce
Just run the attached code.

Include code snippet

import mlx.core as mx
import time
import torch


def mlx_sample():
    x = mx.random.normal([8, 16, 128, 128, 32], dtype=mx.float32)
    weight = mx.random.normal([4, 1, 1, 1, 32], dtype=mx.float32)
    stride = [1, 1, 1]
    padding = [0, 0, 0]
    dilation = [1, 1, 1]
    start = time.time()
    n = 10
    for _ in range(n):
        out = mx.conv_general(x, weight, stride, padding, dilation, stream=mx.gpu)
        mx.eval(out)

    print(f'MLX time: {(time.time() - start) * 1000 / n:0.2f}ms')


def torch_sample():
    x = torch.randn([8, 32, 16, 128, 128], dtype=torch.float32, device='mps')
    weight = torch.randn([4, 32, 1, 1, 1], dtype=torch.float32, device='mps')
    bias = torch.randn([4], dtype=torch.float32, device='mps')
    stride = [1, 1, 1]
    padding = [0, 0, 0]
    dilation = [1, 1, 1]
    start = time.time()
    n = 10
    for _ in range(n):
        out = torch.convolution(x, weight, bias, stride, padding, dilation, False, [0, 0, 0], 1)
        out.max()

    print(f'MPS time: {(time.time() - start) * 1000 / n:0.2f}ms')


def main():
    mlx_sample()
    torch_sample()


if __name__ == '__main__':
    main()

Output:
MLX time: 20.65ms
MPS time: 0.93ms

Expected behavior
At least the same speed as in PyTorch.

Desktop (please complete the following information):

• OS Version: [e.g. MacOS 14.6.1]
• Version [e.g. 0.17.3]
• Tested on M2 Ultra and M1 Pro Max

Start by running the supplied Python benchmark for mlx.core.conv_general and compare its MLX and PyTorch timings on the reported Apple silicon setups. Investigate the conv_general implementation and its GPU execution path, then verify that the benchmark shows performance closer to the PyTorch result.