[upd] operator for benchmark

Context

Many fixes and updates for benchmarked operators (see aidge_benchmark#7)

⚠️ changes made to DataFormat break the export _mem_info file because of enum string call that changes from 'nchw' to 'dformat.nchw'. So export_cpp branch should also be merged

Main changes

• [fix] clone of OperatorTensor by copying dims and MetaOperator forward by forwarding dims
• [add] Scatter and Softplus Python binding
  • [add] Softplus operator generator
• add conversion numpy type / aidge type functions to pybind_Tensor
• [upd] ConstantOfShape add constructor from value and fix some checks
• [add] default value for Pad 'axes' attribute
• [fix] operator Python binding name typos
• Make some formating and add an in_bound() function utils used in LayerNorm
• [upd] make enum to Python standard (uppercase)
• [upd] RandomNormalLike forwardDtype propagation to follow ONNX standard
• [fix] scalar dimension propagation in Add/Sum/Max

Enum

• [upd] change enum for native_enum in Python bindings
• allow to use string instead of enum in some operator constructor
• [add] EnumString function to get enum from string
• [upd] Pybind11 version to 3.0.1 to use native_enum and better support of unique_ptr
• [upd] some operator enum
  • ScatterReduction -> ScatterOp::Reduction
  • BitshiftDirection -> Bitshift_Op::Direction
  • PadMode -> PaddingMode
  • move some operator attributes enum inside the operator
  • change RoundingMode string to snake_case