Add recipe: remove nop_dropout

Context

This MR is part of #83

Problem: ONNX models contain dropout layers with probability=0.0 that act as no-ops during inference, adding unnecessary computational overhead.

Solution: Implement a new recipe to automatically detect and remove these redundant dropout nodes while preserving model functionality.

Required fix: The dropout node implementation should be updated to follow ONNX specification more closely (see issue aidge_core#321).

Modified files

• recipes/nop_dropout.py - New recipe implementation
• recipes/__init__.py - Recipe registration
• unit_tests/onnx_cleaner_models/nop_dropout.py - Test cases
• unit_tests/onnx_cleaner_models/__init__.py - Test integration
• unit_tests/test_onnx_cleaner.py - Test suite update

Detailed major modifications

Recipe Logic:

• Filters dropout nodes with probability=0.0 and simple connections (≤1 children)
• Uses SinglePassGraphMatching for safe node removal
• Preserves complex connections to avoid breaking graph structure

Test Coverage:

• Input/output dropout removal scenarios
• Complex connection preservation
• Non-zero probability preservation

TODO

• Basic implementation
• Test coverage
• Pipeline integration
• Updated attributes handling (training_mode)
• Complex connection handling