Optimize the cat operation on contiguous tensors

[LaurentMazare]

Add a dedicated kernel for efficiently doing the copy in this case.

Benchmarks

Concatenating a (1, 32, 2000, 128) tensor with a (1, 32, 1, 128) one which is typical for a kv-cache operation.

• On cpu (ryzen 2600x), before 9.05ms, after: 7.3ms.
• On cpu (macbook pro, M2 pro 16GB), before 3.8ms, after 1.95ms.
• On gpu (RTX 2080 - 8GB), before 480us, after 290us (this is in kernel blocking mode).

Using a q4k quantized llama 7b model, generating a 1k sequence.

• On gpu (RTX 2080 - 8GB), before 19.9 token/s, after 22.6 token/s.
• On gpu (macbook pro, M2 pro 16GB with metal), before 6.3 token/s, after 9.3 token/s.

Ideally this would use cudaMemcpy2d on the cuda side rather than the current kernel which does a division. Another possibility would be to have specialized kernels for common shapes (e.g. powers of 2 for d2).