Matmul after reshape crashes on Metal

Doing a matrix multiplication after `transpose()` calls / after `permute()` crashes on Metal backend.
Here's a reproduction of the bug:

```rust
use candle_core::{DType, Device, Tensor};

pub fn main() -> candle_core::Result<()> {
let dev = Device::Cpu;
let x = Tensor::zeros(&[1, 8, 52, 1250], DType::F32, &dev)?;
let y = Tensor::zeros(&[1, 8, 1250, 52], DType::F32, &dev)?;
let z = x.matmul(&y)?; // works fine
dbg!(z.shape());

let x = Tensor::zeros(&[1, 8, 1250, 52], DType::F32, &dev)?;
let y = Tensor::zeros(&[1, 8, 52, 1250], DType::F32, &dev)?;
let z = x.matmul(&y)?; // works fine
dbg!(z.shape());

let x = Tensor::zeros(&[1, 1250, 416], DType::F32, &dev)?;
let y = Tensor::zeros(&[1, 1250, 416], DType::F32, &dev)?;

let x = x.reshape((1, 1250, 8, 52))?.transpose(1, 2)?;
let y = y
.reshape((1, 1250, 8, 52))?
.transpose(1, 2)?
.transpose(2, 3)?;

let z = x.matmul(&y)?; // works fine
dbg!(z.shape());

let dev = Device::new_metal(0)?;
let x = Tensor::zeros(&[1, 8, 52, 1250], DType::F32, &dev)?;
let y = Tensor::zeros(&[1, 8, 1250, 52], DType::F32, &dev)?;
let z = x.matmul(&y)?; // works fine
dbg!(z.shape());

let x = Tensor::zeros(&[1, 8, 1250, 52], DType::F32, &dev)?;
let y = Tensor::zeros(&[1, 8, 52, 1250], DType::F32, &dev)?;
let z = x.matmul(&y)?; // works fine
dbg!(z.shape());

let x = Tensor::zeros(&[1, 1250, 416], DType::F32, &dev)?;
let y = Tensor::zeros(&[1, 1250, 416], DType::F32, &dev)?;

let x = x.reshape((1, 1250, 8, 52))?.transpose(1, 2)?;
let y = y
.reshape((1, 1250, 8, 52))?
.transpose(1, 2)?
.transpose(2, 3)?;
// also crashes if you do permute(0, 2, 3, 1)
// also crashes if you create the array with shape (1, 1250, 8, 52) without reshaping

let z = x.matmul(&y)?; // crashes
dbg!(z.shape());

Ok(())
}
```

Start by running the reproduction with Device::new_metal(0) and compare it with the CPU path. Trace Tensor::matmul after reshape, transpose, and permute, focusing on the Metal backend behavior; done means the shown non-contiguous inputs no longer crash and the result matches the working CPU cases.

Written by the indexing model from the issue text.