RDD's Don't cache in some situations with new filegroup reader + new parquet file format

"Test Call rollback_to_instant Procedure with refreshTable" 

Fails if a projection is added to the query plan. The test does not currently fail, because we don't do the project for non-partitioned tables. Adding the projection prevents the rdd from being cached.

Query plans:

without projection, caching works:
{code:java}
== Parsed Logical Plan =='Project ['id]+- SubqueryAlias spark_catalog.default.h0 +- Relation default.h0[_hoodie_commit_time#547,_hoodie_commit_seqno#548,_hoodie_record_key#549,_hoodie_partition_path#550,_hoodie_file_name#551,id#552,name#553,price#554,ts#555L] parquet
== Analyzed Logical Plan ==id: intProject [id#552]+- SubqueryAlias spark_catalog.default.h0 +- Relation default.h0[_hoodie_commit_time#547,_hoodie_commit_seqno#548,_hoodie_record_key#549,_hoodie_partition_path#550,_hoodie_file_name#551,id#552,name#553,price#554,ts#555L] parquet
== Optimized Logical Plan ==InMemoryRelation [id#552], StorageLevel(disk, memory, deserialized, 1 replicas) +- *(1) ColumnarToRow +- FileScan parquet default.h0[id#552] Batched: true, DataFilters: [], Format: Parquet, Location: HoodieFileIndex(1 paths)[file:/private/var/folders/d0/l7mfhzl1661byhh3mbyg5fv00000gn/T/spark-87b3..., PartitionFilters: [], PushedFilters: [], ReadSchema: struct
== Physical Plan ==InMemoryTableScan [id#552] +- InMemoryRelation [id#552], StorageLevel(disk, memory, deserialized, 1 replicas) +- *(1) ColumnarToRow +- FileScan parquet default.h0[id#552] Batched: true, DataFilters: [], Format: Parquet, Location: HoodieFileIndex(1 paths)[file:/private/var/folders/d0/l7mfhzl1661byhh3mbyg5fv00000gn/T/spark-87b3..., PartitionFilters: [], PushedFilters: [], ReadSchema: struct {code}
With projection, no caching:
{code:java}
== Parsed Logical Plan =='Project ['id]+- SubqueryAlias spark_catalog.default.h0 +- Relation default.h0[_hoodie_commit_time#539,_hoodie_commit_seqno#540,_hoodie_record_key#541,_hoodie_partition_path#542,_hoodie_file_name#543,id#544,name#545,price#546,ts#547L] parquet
== Analyzed Logical Plan ==id: intProject [id#544]+- SubqueryAlias spark_catalog.default.h0 +- Relation default.h0[_hoodie_commit_time#539,_hoodie_commit_seqno#540,_hoodie_record_key#541,_hoodie_partition_path#542,_hoodie_file_name#543,id#544,name#545,price#546,ts#547L] parquet
== Optimized Logical Plan ==Project [id#544]+- Relation default.h0[_hoodie_commit_time#539,_hoodie_commit_seqno#540,_hoodie_record_key#541,_hoodie_partition_path#542,_hoodie_file_name#543,id#544,name#545,price#546,ts#547L] parquet
== Physical Plan ==*(1) ColumnarToRow+- FileScan parquet default.h0[id#544] Batched: true, DataFilters: [], Format: Parquet, Location: HoodieFileIndex(1 paths)[file:/private/var/folders/d0/l7mfhzl1661byhh3mbyg5fv00000gn/T/spark-8c60..., PartitionFilters: [], PushedFilters: [], ReadSchema: struct

{code}

## JIRA info

- Link: https://issues.apache.org/jira/browse/HUDI-7162
- Type: Bug
- Epic: https://issues.apache.org/jira/browse/HUDI-6568

Start with the test named “Test Call rollback_to_instant Procedure with refreshTable” and compare its query plans with and without the projection. Trace the new filegroup reader and Parquet scan path to determine why the projected plan does not produce an InMemoryRelation. Done means the projection case preserves RDD caching and the refreshTable rollback test passes.

Written by the indexing model from the issue text.