Add ONNXModifier for optimising the ONNX model before converting for RVC4 execution

[ptoupas]

The ONNXModifier performs a series of onnx modifications that were found to be beneficial in terms of performance on RVC4, without altering the final outputs of the model. Below is the list of optimisations/modifications introduced:

• Substituting division layers for multiplication layers (when possible).
• Substituting subtraction layers for addition layers (when possible).
• Fusing add/mul layers that come right after a conv layer, into a BatchNormalization layer, which is then fused into the conv (when possible).
• Fusing add/mul layers that precede a conv layer into the conv layer by incorporating the addition and multiplication into the conv's weights and biases when possible.
• Removing the Split and Concat layers at the start of the model by reordering the covolutional weights in first conv layer when such transformation is possible.

[kozlov721]

Thanks! LGTM in general, just left a few comments.

[kozlov721]

Could we use graphsurgeon to also simplify the code in onnx_attach_normalization_to_inputs? Or add it to the ONNXModifier

[ptoupas]

Even though I would like the onnx_attach_normalization_to_inputs to be another function of ONNXModifier, I think its better to leave this one outside of the class. The reason is that its the fallback solution in case the optimisations fail or do not produce the correct results compared to the "base" model which is produced by this exact function.

[ptoupas]

But the onnx_attach_normalization_to_inputs alone can be changed to use graphsurgeon instead, yeah.

[aljazkonec1]

Looks good, some comments mostly regarding where and what operations to check.

[klemen1999]

LGTM

[klemen1999]

LGTM