Reduce Python Layer Overhead

[azmyrajab]

This PR speeds up runtime by reducing overhead in python layer, specifically:

• no unnecessary WLS weighting multiplications happen unless user has provided sample weights (it was multiplying expression by 1 implicitly before, which I assumed polars query engine cancelled - but it didn't causing some delay)
• All casting to float64 happens in rust instead of python: it is faster to do there
• refactors code-paths in least_squares.py to minimize duplication (cosmetic)
• documents Maintain input series names, in rust, when a plugin is called within .over() context pola-rs/pyo3-polars#79 which, once resolved, will allow relying fully on rust layer to label struct output for mode="coefficients"