Add max-weight simple cycle detection via message passing

[DhyeyMavani2003]

Implements support for finding maximum weight simple cycles in directed graphs using an iterative path expansion approach. This addresses issue #10320 which requested a PyG solution for computing highest-weight simple directed cycles.

Features:

• find_max_weight_cycle(): Finds the single cycle with highest total weight
• find_all_cycles(): Enumerates all simple cycles with optional filtering

Key capabilities:

• Handles graphs with negative edge weights (e.g., DAGs with reversed edges)
• Ensures simple cycles (no repeated nodes) via visited node tracking
• Memory-efficient with configurable top-k path pruning
• Supports CUDA tensors for GPU acceleration
• Comprehensive test coverage with 15 test cases

Implementation details:

• Uses iterative path expansion with bit masks for visited nodes
• Tracks path states: (current_node, weight, visited_mask, path_history)
• Configurable max_cycle_length and top_k_paths parameters
• Integrates with existing PyG utilities (coalesce, sort_edge_index)

Files added:

• torch_geometric/utils/cycle.py: Core implementation (306 lines)
• test/utils/test_cycle.py: Comprehensive test suite (221 lines)
• examples/cycle_detection.py: Usage examples (194 lines)
• torch_geometric/utils/init.py: Updated exports

Closes #10320