Replicate transform in k

[CongMa13]

Demo code. Will not merge.

In this PR, I want to have multiple threads load same data from input tensor AQ.

AQ

** In VGPR**

How to replicate data?

• Don't create copy in global memory.
• Replicate the view to make multiple addresses map to the same data in global memory

transform:

AQ[16, 2] 
==replicate by 16==> rep_AQ[16, 16, 2] 
==pass_through dim 1 to dim 0; merge dim 2, 0 to dim1 ==>  merge_rep_AQ[16, 32]

Apply replicate and merge transforms on AQ, we can get merge_rep_AQ

Now there are duplicated elements(address) in merge_rep_AQ. We can have multiple threads load the same data with regular tile distribution.