[Feature] Support generating splits with finer granularity than file level

[Zouxxyy]

Motivation

If we have a single Parquet file that is 4GB in size but contains 10 row groups, we can generate 10 splits. This approach will significantly enhance concurrency.

A rough implementation path

• Expand the DataSplit interface to include support for file-level start offset and length.
• Then, implement the ability to split files by offset and length to generate corresponding data splits (for instance, Parquet files can be split further by row groups).
• Finally, ensure that the corresponding readers, such as Parquet, ORC, and Avro readers, support reading rows based on the specified start offset and length + file path.