Aidge cleaner add recipes

This issue track the list of recipes to add to the aidge cleaner modules, ticked recipes are implemented and tested:

• Constant/Shape folding
• Adjust add: tensorrt have issue fusing matmul add based on the order
• Adjust matmul slice: if slice is before matmul it can prevent folding? Not sure why
• EliminateConsecutiveIdempotentOps: idempotent_ops defined by the framework {"Ceil", "Floor", "Round", "Relu", "Reshape"}
• RemoveDeadEnd: Remove op that are not tagged as output but that are at a leaf
• Eliminate Duplicate Initializer: Remove initializer with the same values.
• Eliminate Identity: remove identity nodes
• Eliminate DopOut nodes (should be removed for inference perf)
• Eliminate if with constant condition
• Eliminate_nop_cast: Remove cast if the dtype to cast to is already the input dtype
• nop_concat: Remove concat with one input
• nop dropout: Remove dropout with ratio at 0
• nop expand: Remove exand with shape = 1
• nop flatten: Remove flatten if input is already flatten
• nop monotone argmax: remove operator beofre argmax if it is a monotone operator. two types of monotomne operator, one that depend on the axis (softmax, LogSoftmax) and ones that doesn't depends on the axis (log, exp, sqrt)
• nop pad: Pad with 0s
• nop reshape: Shape = input shape
• nop split: Split to one output
• nop transpose: Tranpose in order
• nop cast: input.dtype = to()
• nop with unit:
  • Remove Add, Or with input tensor 0
  • Remove Mul, And with input tensor 1
  • Remove Sub if input idx = 1 with value = 0
  • Remove Div, Pow if input idx = 1 with value = 1
  • Remove Concat if one of the input is empty
• Shape gather: Replace Shape and Gather by a constant (condition for their constant folding)
• shape op: Replace shape operator by a constant with the shape (condition for their graph constant folding...)
• Slice after shape: Precompute sliced shape
• Unused intializer: Remove initializer connected to nothing (automatically done at aidge import?)
• constant to initializer: Transform constant node to initializer (I there really a gain?)
• Fuse Add bias to conv: Conv followed by Add is fused in Conv bias
• fuse bn conv
• fuse concat reshape: Precompute reshape (constant folding)
• Fuse consecutive concat
• Fuse log -> softmax in LogSoftmax
• Fuse consecutive reduce unsqueeze: If keep dims is false and there is an unsqueeze then set keep dims to true
• Fuse consecutive slices
• Fuse consecutive Squeeze
• Fuse consecutive Transpose
• Fuse consecutive Unsqueeze
• Fuse consecutive Reshape
• Fuse MatlMul Add => Gemm
• Fuse Gelu
• Fuse LayerNorm
• Fuse pad with conv
• Fuse pad with pool
• Fuse QKV
• Fuse transpose into gemm: Transpose weights instead of input