Unmappable dims used for inlining should be leaf domains instead of root domains

[zasdfgbnm]

🐛 Describe the bug

Currently, inlining uses unmappable dims in MaxPosCalculator to avoid inlining into illegal dims. However, the unmappable dims in MaxPosCalculator only contains root domains, which is a mistake and can result wrong CA position being set. See the following example:

TEST_F(NVFuserTest, FusionUnmappableLeaf_CUDA) {
  Fusion fusion;
  FusionGuard fg(&fusion);

  TensorView* tv0 = makeConcreteTensor({4});
  fusion.addInput(tv0);
  auto tv1 = sin(tv0);
  auto tv2 = cos(tv1);
  auto tv3 = exp(tv2);
  auto tv4 = add(tv1, tv3);
  fusion.addOutput(tv4);

  tv3->split(0, 2);
  inlineMost();

  // For old code, use
  // InlinePropagator ip(tv0, -1, ComputeAtMode::MostInlined);
  // MaxRootDomainInfoSpanningTree(tv0).traverse(&ip);

  auto options = at::TensorOptions().dtype(at::kFloat).device(at::kCUDA, 0);

  at::Tensor input = at::randn({4}, options);

  FusionExecutor fe;
  fe.compileFusion(&fusion);
  auto cg_outputs = fe.runFusion({input});

  auto t1 = input.sin();
  at::Tensor output_ref = t1 + t1.cos().exp();

  testValidate(&fusion, cg_outputs, {input}, {output_ref}, __LINE__, __FILE__);
}

Note that this bug is NOT introduced in #2004

Versions

TOT devel

[zasdfgbnm]

Self assign as one of my upcomming refactoring of MaxPosCalculator should fix this. Will not be high priority because I don't see any scheduler triggers such an error.

[zasdfgbnm]

cc @naoyam @csarofeen