Update ONNX cleaner

Context

ONNX cleaner is a new feature of Aidge whihc aim at cleaning/simplifying ONNX graphs. Multiple recipes have been idenitifed and listed in #83.

This MR aim at adding the basis to then add as many reicpes as possible at once.

For this I worked on two recipes which I see as difficult to stress test my implementation of the cleaner, fuseGelu and fuseLayerNorm.

Once the code base is stable I will add unit test on the existing recipes and merge this.

Detailed major modifications

• Update the way recipes are registered, now using a class to specify the API
• Allow to skip recipes using --skip arg:

onnx_cleaner layernorm_pytorch.onnx t.onnx --skip FuseLayerNorm

• Add fuseGelu recipe
• Add fuseLayerNorm recipe
• Add unit tests that generate ONNX to simplify
  • fusegelu
  • fuselayernorm
  • fuse matmul add
  • fuse batchnormalization
  • constant shape folding
  • remove identity
• onnx cleaner now has irversion and opset arguments
• recipes now depends on an opset level
• Update show_recipes option to increase readbility and show supported opset:

TODO

• Allow to skip recipes in the CLI see #57 (closed)
• Recipes should depends on opset (for example Layernorm was introduced at opset 17 so no need to use it for opset <17) see: #82 (closed)
• finish fuseLayerNorm for keras
• test all current recipes
  • fuse matmul
  • remove identity