allocation order inference creating issue in multidevice_matmul tests

[jjsjann123]

https://github.com/NVIDIA/Fuser/pull/2271/files/4e34f1b810aea792b2a188eb2249f352caae32f8#r1607304649

Issue mentioned in above comment.

I think this is falls under the umbrella issue about reduction schedule's handling of allocation domain #2202 , but the specific example doesn't have any permuted inputs, so allocation order inference probably also shouldn't create such complication.

Opening the issue for myself to investigate.