Applied performance optimizations to the generate_pseudo_legal_moves method discovered through automated code evolution using OpenEvolve.

1. Loop consolidation for pawn promotions: Replaced 4 separate yield statements with a single loop over promotion pieces, reducing function call overhead
2. Early bitwise filtering for double pawn moves: Filter by rank before bit shifting ((single_moves & BB_RANK_3) << 8) instead of after, eliminating redundant operations

• Added test_move_generation_performance.py to demonstrate and verify optimizations
• Validates correctness across multiple board positions (starting position, en passant, promotions)
• Benchmarks performance with 100,000 iterations