[Data] Add a target output file size knob for write ops

Description

Description

Ray Data write ops (write_parquet, etc.) currently expose only min_rows_per_file and max_rows_per_file to control output file sizing. Because these are row-count based, output file sizes can vary dramatically when row byte-width varies . This makes it hard to produce evenly-sized output files.

Proposal

Add a size-based knob, e.g. target_file_size / target_block_size (bytes), that Ray Data uses to decide when to roll over to a new output file, estimating rows-per-file from observed average row size. This would sit alongside the existing row-based knobs (which can act as hard bounds).

So api will look something like ds.write_parquet(path, target_file_size="128MiB")

Use case

No response