Add propensity score trimming and partition-specific collinearity checks

[pedrohcgs]

Summary

• Add asymmetric propensity score trimming for control units with pscore >= 0.995
• Add trim_level parameter (default 0.995) to compute_pscore() and compute_pscore_rc() functions
• Add keep_ps indicator to exclude poor counterfactual controls from ATT estimation
• Change upper bound from 1 - 1e-16 to 1 - 1e-6 (matches DRDID)
• Apply trimming to ATT weights but NOT to nuisance parameter estimation
• Add partition-specific collinearity detection with two-stage checking
• Add comprehensive test suite including Monte Carlo coverage test

Trimming Logic

Group	Condition	Included in ATT?
Treated (PA4=1)	Always	✓ Yes
Control (PA4=0)	ps < 0.995	✓ Yes
Control (PA4=0)	ps >= 0.995	✗ No (weight = 0)

Why This Matters

Control units with propensity scores very close to 1 are problematic because:

1. Poor counterfactuals: They are "almost treated" based on covariates
2. Numerical instability: IPW weights explode since ps/(1-ps) → ∞ as ps → 1

Partition-Specific Collinearity

Added check_partition_collinearity() helper that:

• Detects covariates collinear only within specific subgroup comparisons (4 vs 3, 4 vs 2, 4 vs 1)
• Provides informative warnings showing which partitions have issues
• Uses two-stage checking: global collinearity first, then partition-specific

Reference

DRDID package: R/drdid_panel.R, lines 122-143