Slice should clip to end of tensor

[jacobhinkle]

Slice behavior currently does not match numpy/PyTorch when the end_indices extend past the tensor extent:

// Slice with end beyond size of input. This should clip to input, not pad.
TEST_F(NVFuserTest, FusionResizeSlice6_CUDA) {
  Fusion fusion;
  FusionGuard fg(&fusion);

  std::vector<int64_t> shape({9});

  // concrete shapes to avoid dynamic Fusion
  auto tv0 = makeConcreteTensor(shape);
  fusion.addInput(tv0);

  auto tv1 = slice(tv0, {{fusion.zeroVal(), IrBuilder::create<Int>(11)}});
  fusion.addOutput(tv1);

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

  auto t0 = at::randn(shape, options);
  std::vector<c10::IValue> aten_inputs({t0});

  FusionExecutor fe;
  fe.compileFusion(&fusion, aten_inputs);
  auto cg_outputs = fe.runFusion(aten_inputs);

  auto ref = t0.index({at::indexing::Slice(0, 11)});

  testValidate(&fusion, cg_outputs, aten_inputs, {ref}, __LINE__, __FILE__);
  // C++ exception with description "The size of tensor a (9) must match the
  // size of tensor b (11) at non-singleton dimension 0 Exception raised from
  // infer_size_impl at /opt/pytorch/pytorch/aten/src/ATen/ExpandUtils.cpp:31
}

In these cases, we should clip to the end of the tensor. That is, we should implicitly use min(extent, end_index) instead of index itself. As discussed in #397 this means we should not concretize slice ops prematurely unless we know their input extents.