Support symbolic -1 output dimensions in reshape

[jacobhinkle]

As of #24 and #215 there is support for scheduling and running Fusions containing reshape ops whose output shape is a list of Scalars. These scalars are typically the extents of the corresponding reshaped dimensions. If -1 is passed, it should be replaced by an integer that preserves the number of elements in the input. Currently, we are able to handle dynamic values of -1 during concretization, but scheduling fails while getting heuristics.

  std::unique_ptr<Fusion> fusion_ptr = std::make_unique<Fusion>();
  Fusion& fusion = *fusion_ptr.get();
  FusionGuard fg(&fusion);

  TensorView* tv0 = makeSymbolicTensor(4);
  fusion.addInput(tv0);

  auto tv1 = add(tv0, tv0);
  auto tv2 = reshape(tv1, {tv1->axis(0)->extent(), tv1->axis(2)->extent(), IrBuilder::create<Int>(-1)});
  auto tv3 = add(tv2, tv2);
  fusion.addOutput(tv3);

  FusionExecutorCache fusion_executor_cache(std::move(fusion_ptr));


  auto options = at::TensorOptions().dtype(at::kFloat).device(at::kCUDA, 0);
  at::Tensor at_x = at::randn({2,3,4,5}, options);

  auto outputs = fusion_executor_cache.runFusionWithInputs({at_x});
//C++ exception with description "extent_int >= 0 INTERNAL ASSERT FAILED at 
// "/opt/pytorch/nvfuser/csrc/scheduler/registry.cpp":807, please report a bug to PyTorch. 
// Unexpected size of axis: -1                                                                                                  
// Exception raised from getTensorIndexType at
// /opt/pytorch/nvfuser/csrc/scheduler/registry.cpp:807