[Fix][CUDA] Allocate complete words for shared sub-byte buffers

[LittlehamsterXu]

Description

CUDA codegen packs int1, int4, and uint4 values in shared memory using 32-bit words. The static shared-memory allocation size was previously computed with floor division.

For shapes that are smaller than one packing word, this could generate a zero-sized __shared__ array. For non-divisible shapes, it could also under-allocate the final partial word.

This patch uses ceil division so that every partial final word is allocated. It also adds regression coverage for int1, int4, and uint4 shapes.

Fixes #20273

Testing

• pytest -q tests/python/tirx/codegen/test_codegen_cuda.py -k subbyte_shared_alloc_uses_ceil_div
• pytest -q tests/python/tirx/codegen/test_codegen_cuda.py
• ruff check tests/python/tirx/codegen/test_codegen_cuda.py
• git diff --check