Check Tensor.dims equals the rank of the corresponding TensorView.

[wujingyue]

This is a follow-up to #3114. A bug could have been surfaced earlier if we had checked the equality.

To compare a Tensor and its TensorView, you'd have to construct the cpp Fusion by calling fd.execute. Then, for each tensor, get its TensorView via getFusionState(size_t index), then compare their dimensions.

Originally posted by @rdspring1 in #3114 (review)

[wujingyue]

Is it possible to check while building the fusion IR? Most errors (shown as exceptions) I encountered happen during

Fuser/csrc/python_frontend/fusion_definition.cpp

Line 97 in 6c3f264

buildFusionIr(preschedFusion());

; they don't even reach verifyTensorDimensions.

(Sorry, I should have noticed the difference earlier)

[wujingyue]

Regardless of my last comment, the verification didn't seem to kick in for the following example.

git checkout wjy/3117
_bn && NVFUSER_DUMP=fusion_ir_original pytest tests/python/test_matmul.py -s -k 2d_x_3d

Tensor(index=2, ndim=2)
Inputs:
  T0_g_float[ iS0{2}, iS1{3} ]
  T1_g_float[ iS2{7}, iS3{3}, iS4{5} ]
Outputs:
  T2_g_float[ iS5{7}, iS6{2}, iS7{5}, rS8{3} ]

%kernel_math {
T2_g_float[ iS5{7}, iS6{2}, iS7{5}, rS8{3} ]
   = matmul(T0_g_float[ iS0{2}, iS1{3} ],
            T1_g_float[ iS2{7}, iS3{3}, iS4{5} ])
} // %kernel_math

Tensor(index=2, ndim=2)

c is 2D and T2 is 3D. However, verification didn't flag that.