Replace 50% downsampling with larger L0 penalty for better accuracy

[MaxGhenis]

Problem

Currently, CPS datasets from 2019-2023 use frac = 0.5 to randomly downsample to 50% of observations before reweighting. This reduces memory usage but throws away half the data randomly, potentially losing important observations.

Proposed Solution

Remove the 50% downsampling and instead use a larger L0 penalty in the reweighting step. The L0 penalty encourages sparsity by setting some weights to exactly zero, achieving similar computational benefits but in a learned way that preserves important observations.

Benefits

• Better accuracy: The model decides which observations to effectively zero out based on their contribution to targets
• Same memory footprint: Similar sparsity level (≈50% zeros) but chosen optimally
• More principled: Uses optimization rather than random sampling

Implementation

1. Remove frac = 0.5 from CPS_2019 through CPS_2023 classes
2. Increase the L0 penalty parameter in the reweighting function to achieve approximately 50% sparsity
3. Test to find the L0 value that gives similar sparsity to current approach

Related to #391 (L0 exploration)