Rewrite ParallelDimensionMap with expr simplifier

[zasdfgbnm]

I am seeing trivial thread predicate in one matmul schedule #87, and that trivial thread predicate causes a 10% slowdown. So I start my work on removing that predicate. My first step is a rewrite of ParallelDimensionMap. The new code is much shorter in lines of code, and should have better result (for example, in FusionParallelDimensionMap3_CUDA, blockDim.x becomes 20).

This PR uses the following equation to calculate the extent of a parallel dim:

parallel_dim = simplifyExpr(max(extent1, extent2, ..., extentN));

Future work:

• Simplify trivial predicate in Implement with split merge mmigdal-nv/Fuser#1, what we need to do is:
  • When simplifying all expressions, we should assume parallel_index < parallel_dim in additional to

    Fuser/csrc/expr_simplifier.cpp

    Line 152 in 3e69f69

    axioms.push_back(SimplifyingIrBuilder::ltExpr(pidx, pdim));

  • Omit the predicate not only when isExact(ptype), but also when the loop extent equals parallel_dim.
• Simplify the following to use the extent computed in this PR:
  

  Fuser/csrc/executor.cpp

  Line 660 in 548d5d2

  for (auto& entry : simplified_parallel_iter_extents) {

[zasdfgbnm]

at::cuda::warp_size(); was added to support ROCm, which we no longer care anymore.

[naoyam]

Nice cleanup!

[zasdfgbnm]

!build

[naoyam]

LGTM