Llama Support

[Victor-Jung]

This PR adds support for quantizing Llama-flavour transformers. It provides new operators, related passes as well as various fixes.

Added

• IntegerRMSNorm operator and integerization pass
• IntegerTrueDiv operator and integerization pass
• PactifyPass: a pass to replace all Conv, Linear and Relu node into their PACT equivalent with the given configuration

  Changed

• ApproximateSiLUPass now approximates SiLU with HardSwish instead of GELU
• IntegerHardSwish now have a symbolic ONNX node representation
• Passes that were using PACT_symbolic_trace as a default tracer can now be initialized with a custom symbolic trace

  Fixed

• Patch the optimization configuration of export_net function for onnxruntime version above 1.16
• Fix PACTIntegerConcat
• Fix default args of PACTIntegerConcat nodes inserted by ConcatTreeReplacementPass

  Removed

• Remove some unused exports