Make user-defined BulkInsertPartitioner fit write path API

this existing logic is problematic due to we can’t enforce user’s partitioner to return JavaRDD, this potentially breaks.

{code:java}
BulkInsertPartitioner partitioner = userDefinedBulkInsertPartitioner.isPresent()
? userDefinedBulkInsertPartitioner.get()
: BulkInsertInternalPartitionerFactory.get(config.getBulkInsertSortMode());
repartitionedRecords = (JavaRDD>) partitioner.repartitionRecords(dedupedRecords, parallelism);
{code}

The factory is used only in spark for now. So, we expect JavaRDD or HoodieData. The API can be made explicit about the constraint.

## JIRA info

- Link: https://issues.apache.org/jira/browse/HUDI-3463
- Type: Improvement

Start with the write-path code shown around BulkInsertPartitioner, BulkInsertInternalPartitionerFactory, and the repartitionRecords call. Trace how JavaRDD and HoodieData are handled, then define the API constraint for user-defined partitioners. Done means the write path accepts only the supported return type without relying on the shown cast.

Written by the indexing model from the issue text.