Enhance dropout

Context

This Merge Request adds full support for the ONNX Dropout operator (opset ≥ 12), based on the ONNX specification. It introduces three key features:

1. training_mode input (optional)

   • A boolean scalar input that controls whether dropout is applied (True) or bypassed (False).
   • If omitted, it defaults to inference mode (True), meaning dropout is applied.
   • Enables dynamic control of training vs. inference behavior at runtime.

2. mask output (optional second output)

   • A boolean tensor of the same shape as the input.
   • Indicates which elements were retained (True) or dropped (False) during dropout.
   • Useful for debugging, backpropagation, or custom downstream logic.

3. `seed attribute (optional)

   • A fixed integer value used to seed the random number generator for dropout.
   • Ensures reproducibility of the dropout mask across runs when provided.
   • If omitted, a random seed is used (non-deterministic behavior).

resolves issue aidge_core#321

Implementation Details

• Added parsing and validation for the optional training_mode input:

  • Defaults to True (training mode) if not provided
  • Disables dropout logic when training_mode=False

• Implemented generation of the mask tensor when the second output is requested:

  • Mask is a boolean tensor indicating retained elements
  • Same mask is used to compute the final output

• Supported the optional seed attribute for deterministic behavior:

  • When provided, it seeds the dropout RNG
  • Guarantees reproducible masks across runs

• Updated operator registry and unit tests accordingly

TODO

• training_mode attribute
• mask output
• seed attribute
• unit_tests