Fix the batch plan for a bulk CSV load

An AI's ETL step read an 80 million row CSV into memory with one pd.read_csv call, then inserted it one row at a time. It ran out of memory, and on smaller files it took hours. The rewrite streams the file in chunks and sends each chunk to Postgres in multi-row batches. You plan the statement sizes before running it.

An assistant wrote plan_bulk_load(total_rows, chunk_size, batch_size). It is in the editor. It has a bug. Find it and fix it.

Input and output:

• total_rows: rows in the file.
• chunk_size: the most rows read into memory at a time.
• batch_size: the most rows sent in one insert statement.
• Return the number of rows in each insert statement, in order.

Rules the tests check:

• If any argument is 0 or less, return [].
• Read the file in chunks of chunk_size rows. The last chunk holds whatever is left, so every row is read exactly once.
• Split each chunk into batches of batch_size rows. The last batch of a chunk holds whatever is left of that chunk. A batch never mixes rows from two chunks.
• The sizes add up to total_rows.