Fix #3607

[naoyam]

Fixes #3607.

In the fusion reported in #3607, there's a proposed fusion segment as shown below:

**Segmenter** Considering fusion:
T14_g_double[iS27{2}, rS28{24000}](Avg),
T15_l_double[iS29{2}, rS30{24000}](Var),
T16_l_nvfuser_index_t[iS31{2}, rS32{24000}](Count)
 = Welford ( T2_g_double[iS3{2}, iS4{24000}](Avg),
  allreduce = false )
d59 = (double)(24000);
d61 = double(1) * d59;
d65 = (double)(0);
d67 = d61 - d65;
d69 = (double)(0);
b71 = d67 >= d69;
d73 = (double)(0);
d75 = where(b71, d67, d73);
d80 = reciprocal(d75);
T17_l_double[iS33{2}]
   = T15_l_double[iS29{2}, rS30{24000}]
   * d80;
T23_l_double[iS44{2}, bS45{1}]
   = broadcast( T17_l_double[iS33{2}] )
T25_l_double[iS48{2}, bS49{1}]
   = T23_l_double[iS44{2}, bS45{1}]
   + double(1.0000000000000001e-05);
T26_g_double[iS50{2}, bS51{1}]
   = rsqrt(T25_l_double[iS48{2}, bS49{1}]);
T27_l_double[iS52{2}, bS53{1}]
   = Set( T26_g_double[iS50{2}, bS51{1}], cache_op=Streaming )
T28_g_double[iS54{2}, bS55{1 ex 24000}] = expand( T27_l_double[iS52{2}, bS53{1}], {2, 24000} )
T32_g_double[iS62{2}, iS63{24000}]
   = T28_g_double[iS54{2}, bS55{1 ex 24000}]
   * T13_g_double[iS25{2}, iS26{24000}];
T37_l_double[iS72{2}, iS73{24000}]
   = -T32_g_double[iS62{2}, iS63{24000}];
T38_g_double[iS74{2}, rS75{24000}]
   = reduction( T37_l_double[iS72{2}, iS73{24000}], op = add, initial value = double(0), allreduce = false )

When canScheduleCompileTime of the transpose scheduler is called, FindAllMappedDims is used with T32 as the reference. Since its innermost dimension is not connected with T15 due to the reduction, propagateSibling hits the assertion at https://github.com/NVIDIA/Fuser/blob/main/csrc/scheduler/utils.cpp#L1476.

This PR avoids the assertion by checking the existence of reduction first. We may also want to remove the assertion.

[naoyam]

!test

[naoyam]

I moved this here because because it uses getInputsOutputsWithInnerDim, which uses FindAllMappedDims, which asserts the existence of mappings between sibling tensors.

https://github.com/NVIDIA/Fuser/blob/main/csrc/scheduler/utils.cpp#L1476

I'm not entirely sure why we should have an assertion there. If the from tensor has a mapping to nullptr, which is not asserted, but propagating that to a sibling is asserted. I think we can just remove the assertion, but not 100% sure.

Instead, we shouldn't do this level of analysis until the fusion is confirmed to consist of only pointwise ops with no reduction. This should avoid the issue and also would help reduce the segmentation cost.

[t-vi]

Thank you for this, @naoyam !
I can confirm that this unblock Lightning-AI/lightning-thunder#1560 .