feat(pruning): Implement Lottery Ticket Hypothesis pruning

[noahgift]

Summary

Implement Lottery Ticket Hypothesis (LTH) pruning based on the foundational paper by Frankle & Carlin (2018).

Paper: https://arxiv.org/abs/1803.03635
Title: The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks

Background

The Lottery Ticket Hypothesis proposes that dense neural networks contain sparse subnetworks ("winning tickets") that:

• Are typically 10-20% of original network size
• Achieve comparable test accuracy when trained from scratch
• Train faster and achieve higher accuracy than full networks above minimum threshold
• Depend critically on their initialization weights

Implementation Tasks

• Add LotteryTicketPruner struct to src/pruning/mod.rs
• Implement iterative magnitude pruning with weight rewinding
• Add winning ticket identification algorithm
• Support for rewinding to initialization (IMP - Iterative Magnitude Pruning)
• Add late rewinding variant (rewind to early training, not init)
• Benchmark against existing magnitude pruning

API Design

use aprender::pruning::{LotteryTicketPruner, RewindStrategy};

let pruner = LotteryTicketPruner::builder()
    .target_sparsity(0.9)           // 90% sparse (10% weights remaining)
    .pruning_rounds(10)              // Iterative pruning rounds
    .rewind_strategy(RewindStrategy::Init)  // Rewind to initialization
    .build();

let winning_ticket = pruner.find_ticket(&model, &train_data)?;

Acceptance Criteria

• LTH pruner achieves <20% parameter count with <1% accuracy drop on MNIST
• Winning tickets train faster than dense baseline
• Property tests for weight rewinding correctness
• Book chapter: examples/lottery-ticket-pruning.md
• Benchmark comparison with magnitude pruning

References

• Original Paper: Frankle & Carbin (2018) - https://arxiv.org/abs/1803.03635
• Stabilizing LTH: https://arxiv.org/abs/1903.01611
• Linear Mode Connectivity: https://arxiv.org/abs/1912.05671