[Wave] Fixes to extend attention

[harsh-nod]

This PR adds the following fixes to extend attention:

1. Fixes the way the block indices were being read for the v matrix in the first loop. Because of the layout differences between the first and second mma, we need to read block indices with a different layout and correspondingly elements per thread when loading from the kv cache.
2. Moves the qk-transpose scaling inside the kernel for accuracy reasons.
3. Moves reuse shared allocs pass after minimize global loads.
4. Minor compiler fix regarding use of Sequence vs list

[raikonenfnu]

Overall looks good, just that one quick question