Refactor Sun-Abraham to use true saturated regression approach

This addresses code review feedback by implementing the methodology
from Sun & Abraham (2021) using a saturated TWFE regression instead
of 2x2 DiD comparisons.

Key changes:
- Implement _fit_saturated_regression() with cohort × relative-time
  interaction dummies and two-way within transformation
- Add proper covariate support through the regression framework
- Use cluster-robust standard errors from regression output
- Compute IW effects using delta method with full vcov matrix
- Switch to pairs bootstrap (removed unused bootstrap_weights param)
- Fix linting issues (unused imports, dead code, bare exceptions)
- Optimize DataFrame operations to avoid fragmentation warnings
- Update tests to reflect API changes (29 tests pass)

The new implementation matches the academic literature exactly:
1. Saturated TWFE: Y_it = α_i + λ_t + Σ_g Σ_e δ_{g,e} D_{g,e,it} + X'γ + ε
2. Cohort weights: w_{g,e} = n_g / Σ_g' n_g'
3. IW aggregation: β_e = Σ_g w_{g,e} × δ_{g,e}

Author: claude