[WS] add pingpong pass

[manman-ren]

Initial preliminary implementation for PingPong across the two consumer warp groups. This assumes the consumer warp groups execute the same code, and the pass will try to figure out cuda core region vs. tensor core region and insert named barriers to synchronize across the consumer warp groups. When one consumer warp group is in cuda core, the other should be in tensor core.

The pass is guarded with env variable ENABLE_PINGPONG.

[manman-ren]

@bertmaher This is the initial implementation. The logic is kind of similar to AMD pingpong but AMD pingpong does slicing/clustering/conditional barrier across two waves (two waves execute the same code). It will be great if you can try it out on small-K GEMMs.