[example] Linalg to XeGPU fused attention implementation.

[charithaintc]

This example demonstrate how to optimize standard attention kerel written in linalg level into the fused attention kernel that can be run gpu.

Main steps involved:

1. Generate standard attention payload on 4d tensors (batch x head x ctx_len x d_head)
2. Tile and fuse the outer parallel dims (batch and head)
3. Vectorize/Bufferize
4. Use transform extensions to generate the inner tiled reduction loop (Until we have a better solution).
5. Distribute to GPU workgroups.
6. Set xegpu layouts and lower to binary.

Currently this depends on a fix for : #147