[BugFix][Arith] Reject cyclic floormod normalization in bijective maps

[SiriusNEO]

Summary

• Prevent DetectIterMap(..., Bijective) from silently normalizing a shifted modulo expression to an unshifted one when the offset has a non-zero modulo phase.
• Reject the unsupported cyclic rotation instead of returning an IterMap that InverseAffineIterMap would invert incorrectly.
• Preserve existing behavior for offsets divisible by the modulus and for padding-aware non-bijective simplification.

Related to tile-ai/tilelang#2948 and tile-ai/tilelang#3090.

Root cause

VisitExpr_(AddNode) stores the offset in IterSumExpr::base. SplitFloorModConst computes that phase as pair.second, but the bijective path has no padded IterMark carrying it and returns only pair.first. The resulting normalized map is therefore semantically different:

(x + 1) % 128  ->  x

InverseAffineIterMap then sees an unshifted split and produces x = output.

Validation

Before the fix, the regression test failed because the cyclic shift produced one accepted bijective index.

After the fix:

• cmake --build build -j 32
• PYTHONPATH=$PWD/python TVM_LIBRARY_PATH=$PWD/build python3 -m pytest -q -p no:tvm.testing.plugin tests/python/arith/test_arith_iter_affine_map.py — 40 passed
• python3 -m ruff check tests/python/arith/test_arith_iter_affine_map.py
• python3 -m ruff format --check tests/python/arith/test_arith_iter_affine_map.py
• clang-format --dry-run --Werror src/arith/iter_affine_map.cc
• git diff --check

The repository-wide ASF-header pre-commit hook reports 11 pre-existing missing headers outside this change; all relevant formatting and lint hooks pass.