Add unified and configurable null handling

[ryankert01]

Related Issues

Closes #765

Changes

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Why

How

• Unify null value handling between batch and streaming Parquet/Arrow readers by introducing a configurable NullHandling enum (FillZero | Reject)
• Previously batch mode silently coerced nulls to 0.0 while streaming mode threw a runtime error — both paths now use the same handle_float64_nulls() helper
• Defaults to FillZero for backward compatibility; Reject returns a clear error with guidance
• Threads the policy through PipelineConfig, PyO3 bindings, and the Python QuantumDataLoader builder (.null_handling("fill_zero" | "reject"))

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[viiccwen]

overall lg, but need to update docs, and left some comments.

[viiccwen]

Array is not used in this file.

[viiccwen]

currently we specify read_* ( read_parquet(...) ) to use NullHandling::FillZero, should we adopt same strategy like *Readers (ParquetReader)?

like this?

pub fn read_parquet<P: AsRef<Path>>(path: P) -> Result<Vec<f64>>;
// to
pub fn read_parquet<P: AsRef<Path>>(
    path: P,
    null_handling: Option<NullHandling>,
) -> Result<Vec<f64>>;