Add admm_optimization.rs demonstrating distributed and federated learning
applications of the Alternating Direction Method of Multipliers.

Examples (~620 lines, 4 comprehensive demos):

1. **Distributed Lasso Regression**
   - Consensus form for sparse regression
   - Separates smooth (least squares) and non-smooth (L1) parts
   - Demonstrates adaptive ρ parameter adjustment
   - Recovers sparse coefficients with < 0.1 RMSE

2. **Consensus Optimization (Federated Learning)**
   - Simulates 3 distributed workers with local data
   - ADMM enforces global consensus across workers
   - Demonstrates federated averaging
   - Workers converge to global solution

3. **Quadratic Programming with ADMM**
   - Box-constrained QP: minimize ½xᵀQx + cᵀx s.t. x ≥ 0
   - Consensus form separates objective from constraints
   - Adaptive penalty parameter for faster convergence

4. **ADMM vs FISTA Comparison**
   - Side-by-side on same Lasso problem
   - Shows tradeoffs: ADMM for distributed, FISTA for centralized
   - Performance metrics and convergence analysis
   - Guidance on when to use each method

Key Features Demonstrated:
- Adaptive ρ adjustment (Boyd et al. 2011)
- Consensus constraints for distributed optimization
- Separation of smooth/non-smooth objectives
- Proximal operators (soft-thresholding)
- Cholesky solver for x-subproblems

Applications:
- Distributed data across workers
- Federated learning scenarios
- Model parallelism
- Constrained optimization via consensus

Follows pattern from batch_optimization.rs and convex_optimization.rs
with detailed mathematical background and practical ML applications.

Related: ADMM implementation (commit 9f30ffe)