[FEA] Investigate why padding is needed for input buffers to compression/decompression kernels

After `rmm` removed memory padding (https://github.com/rapidsai/rmm/pull/1278), some places in libcudf got into trouble with out-of-bound memory access. These bugs can be fixed by simply padding the memory buffer back (https://github.com/rapidsai/cudf/pull/13586) but that is not an optimal, long-term solution. The permanent solution should be discovering why such padding is needed and whether we can avoid such padding altogether.

We should investigate further the issue to find a permanent solution.