preseg optimization pass - optimization for cat

[jjsjann123]

Adding pre-segmentation optimization pass that optimize cat (PadOp + CatOp) to avoid multiple kernels.

1. The mutation tries to push PadOp further to its producer as described below. The goal is to either have the pad op directly on inputs to fusion, or to a point where segmentation at pad leaves a no-op fusion segment before pad.

// The concept is that the following program:
//   tv1 = UnaryOp(tv0)
//   tv2 = PadOp(tv1)
// can be replayed as the program below:
//   tv0_padded = PadOp(tv0)
//   tv2_new = UnaryOp(tv0_padded)

For details on the propagation logic, please see code comment
Note [ PadOp Propagation Rule ]
Note [ Handling TV with Multiple Uses via Frontier ]

2. As a result of moving the PadOp, we also need to replace the CatOp with a series of binary add, since its inputs are no longer directly produced by PadOp.

[jjsjann123]

!build

[jjsjann123]

!build

[jjsjann123]

!build

[jjsjann123]

!build --pybench

[jjsjann123]

!build --pybench

[jjsjann123]

!build

[jjsjann123]

Don't see any odd failure in pybench. I'll merge this PR after CI clears.

[jjsjann123]

!build

[jjsjann123]

failures are related to distributed stuff. Don't seem to related but I'm gonna try merging main again.

[jjsjann123]

!build

[jjsjann123]

!build

[jjsjann123]

!build

[jjsjann123]

!build

[jjsjann123]

!build

[jjsjann123]

!build

[jjsjann123]

!build