Wasteful to pull all primary key columns in query of append mode

[Rachelint]

Describe This Problem

We pull projected + pk columns when querying in both append and overwrite mode.
However pulling pk columns is totally unnecessary in append mode.

Proposal

Just pulling the projected columns in query of append mode.

The record batch reading steps will be divided to following threes:

• Pulling the ArrowRecordBatch.
• Converting it to FetchingRecordBatch, somethings like filling not exist column a null/default value will be done here.
  FetchingRecordBatch will be used in ChainIterator/ MergeIterator, and it may include not only the projected columns but the primary key columns for dedupping or else.
• Prune to RecordBatch, as saying above, FetchingRecordBatch can include not only projected columns, prune the non-projecteds here.

Main changes:

• Don't pass the ProjectedSchema including too many informations to where building the inner RecordBatchStream in ChainIterator/ MergeIterator. Instead, refactor the RowProjector to just include the needed informations and pass it to(mainly ScanRequest and SstReadOptions).
• Refactor RecordBatchWithKey to FetchingRecordBatch, the main difference is that FetchingRecordBatch can include primary_keys_indexes or not. Actually, FetchingRecordBatch should not include primary_keys_indexes anymore, but it is hard to remove it completely, so maybe we can delay it to later prs.

Additional Context

No response