Draft: Add InstanceNorm operator for 3DUnet support

This MR is linked to aidge#308 and aims to add the missing operator InstanceNormalization (InstanceNorm) for 3D-Unet support.

InstanceNorm - Core Framework Implementation

Core Framework Files

• InstanceNorm.hpp (202 lines) - Main operator header with epsilon attribute and comprehensive mathematical documentation
• InstanceNorm.cpp (81 lines) - Operator implementation with constructor, dimension forwarding, and factory function
• pybind_InstanceNorm.cpp (128 lines) - Python bindings with LaTeX documentation and parameter defaults
• Test_InstanceNorm_Op.cpp (169 lines) - Unit tests for attributes, dimension handling, and functionality

Core Operator Architecture

Header Implementation (InstanceNorm.hpp)

• Attribute System: StaticAttributes<InstanceNormAttr> with single Epsilon parameter
• Class Definition: InstanceNorm_Op inherits from OperatorTensorWithImpl<InstanceNorm_Op>
• Documentation: Extensive LaTeX formulations for forward/backward passes
• API Design: Epsilon getter and factory function with learnable parameters

Source Implementation (InstanceNorm.cpp)

• Constructor: Epsilon initialization with three input categories (Data, Param, Param)
• Dimension Forwarding: Multi-dimensional tensor shape inference and parameter validation
• Factory Function: Creates nodes with attached scale/bias parameters via addProducer()

Python Bindings (pybind_InstanceNorm.cpp)

• Class Binding: Python class with attribute access and comprehensive docstrings
• Function Binding: Factory with defaults (nb_features, epsilon=1e-5, name="")
• Registration: Automatic operator registration for Python interface

Unit Testing (Test_InstanceNorm_Op.cpp)

• Coverage: 3D/4D tensors, constructors, attributes, factory function, and I/O names
• Validation: Epsilon handling, dimension forwarding, and parameter attachment

Key Implementation Details

• Single Attribute: Epsilon (float, default 1e-5) for numerical stability
• Input Structure: data_input, scale (γ), bias (β) - latter two are learnable per-channel parameters
• Normalization: Per-instance, per-channel across spatial dimensions
• Mathematical Support: Complete forward/backward formulations with gradient computation

External References

• InstanceNorm paper: The Missing Ingredient for Fast Stylization https://arxiv.org/pdf/1607.08022
• Pytorch documentation: https://docs.pytorch.org/docs/stable/generated/torch.nn.InstanceNorm3d.html
• ONNX documentation: https://onnx.ai/onnx/operators/onnx__InstanceNormalization.html