De couple use_dask from CSV parser and make it a property of parallelizer

[gsaluja9]

[ad-claw000]

This issue requests removing the use_dask parameter from CSVParser and its subclasses, moving it instead to be a property of Parallelizer (and by extension ParallelQuery, ParallelLoader, etc.) to better separate data parsing from execution orchestration.

Here are a few possible approaches to implement this:

1. Always load via Pandas in CSVParser; convert to Dask in Parallelizer

   • Pros: Completely strips all Dask imports and logic out of CSVParser, simplifying the class.
   • Cons: Forces the entire dataset into memory via a Pandas DataFrame before Parallelizer converts it using dd.from_pandas(). We lose Dask's lazy out-of-core file loading benefits.

2. Move Dask file-loading logic directly into DaskManager

   • Pros: Maintains out-of-core memory efficiency. DaskManager.run can check if generator has a .filename attribute and use dask.dataframe.read_csv to lazily chunk the file, falling back to dd.from_pandas(generator.df) for non-file generators.
   • Cons: Slightly couples DaskManager to the knowledge that file-based generators use CSV format, though DaskManager already heavily relies on slicing and re-instantiating generator.__class__.

3. Extract data loading into a DataLoader interface

   • Pros: Cleanest architectural separation. Generators implement a get_df(dask=False) or similar method, allowing Parallelizer to dictate how data should be supplied.
   • Cons: A larger refactor that might introduce more abstractions than strictly necessary to fix this issue.

Recommendation:
I recommend Approach 2. We can move the existing dask.dataframe.read_csv blocksize calculation from aperturedb/CSVParser.py directly into aperturedb/DaskManager.py. Parallelizer (in aperturedb/Parallelizer.py) will gain a use_dask property, initialized to False. ParallelQuery will set this property and delegate to DaskManager, which will handle chunking the file intelligently. This achieves the requested decoupling without losing performance.