[Core] Fix PredicateVector zero detection

[migarci2]

Summary

• keep complete predicate masks instead of narrowing them to bool
• include every byte when the last storage word is exactly full
• cover single-word, partial-word, full-word, and padding-bit cases

Testing

• compiled test/unit/core/predicate_vector.cu with NVCC 11.2
• ran PredicateVector.* on an RTX 3090: 3 tests passed
• verified the standalone reproduction fails before the change and passes after it
• git diff --check

The machine has CUDA 11.2, so the unrelated complex.cu part of the aggregate core target cannot build against CUTLASS 4.8. The affected test source compiled and ran successfully.

Contribution disclosure

AI assistance was used for repository research, implementation, independent review, and running the checks above.

Fixes #3512