test_sdpa_loop_split errors out with dynamic shape

[Priya2698]

PR #3807 switches tests exercising DID loop split to use dynamic shapes. The sdpa_loop_split test case currently does not work with dynamic shapes indicating a potential bug there around DID loop split.

CC: @wujingyue

[wujingyue]

FWIW, I added a repro to https://github.com/NVIDIA/Fuser/tree/bug3817

[Priya2698]

I print out some values in the expression evaluator and it seems the inferred values for the shapes are incorrect:

2: StructHandle<Tensor<float, 3, 4>>{data=0, logical_size=2 6 1024, logical_stride=3072 1024 1, alloc_size=1 2 3 1024, alloc_stride=3072 3072 1024 1}

The above is log_sumexp. This is not the case for all tensors.
I am not certain if this is due to the logic of DID loop split, could be that symbolic shapes are not correctly inferred for DID parallelized tvs. I need to debug further to verify the latter.

[wujingyue]

I vaguely remember @csarofeen found a bug with allocation of symbolic shapes when working on #3826. Could be related...