Fix cudf::hash_partition kernel launch error with decimal128 types

[davidwendt]

Description

Fixes cudf::hash_partition error when using decimal128 column types. The internal optimized path, copy_block_partitions, uses shared-memory for copying fixed-width type column elements. For int128_t type, the shared-memory needed (~64KB) is larger than the maximum size (~48KB) allowed causing a kernel launch failure.

The optimized path is now restricted to only fixed-width types int64_t and below. The int128_t column types will fall through to the gather-map pattern instead. Accommodating this type in the existing copy-block implementation would likely penalize the performance of the other fixed-width types. If the new implementation becomes insufficient, we could explore a special optimized path in the future for the single type int128_t.

An existing gtest for fixed-point types was updated to include a decimal128 column to catch this kind of error in the future.

Closes #12852

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[davidwendt]

@revans2 @firestarman Could you test out this fix with the Spark testcase from #12852 ?

[firestarman]

@davidwendt This PR can fix #12852. Thanks a lot.

[davidwendt]

/merge