[Experimental] Switch to new onnx-passes frontend

[fpjentzsch]

Attempt to switch all example/regression test flows to the new onnx-passes frontend introduced here: #116

Not modified yet:

• ResNet-50 (due for replacement)
• (end2end) unit tests

Depends on / includes: #116

[iksnagreb]

I'd recommend to keep the old step_streamline as a fallback for now, there are at least two transformations not covered by the new frontend at the moment: Moving element-wise operations (Add, Mul, ...) past padding and some equivalent/generalization of RemoveCNVtoFCFlatten (which ideally should be considered streamlining and not part of hardware conversion/post-processing), meaning without the old steps there might be scales stuck in front of each Pad operator and in front to the first linear layer following the convolutions.

[fpjentzsch]

The commit history of this PR is now pretty messed up, so it will definitely need to be re-based or squashed.

This PR is still represents an aspirational goal: If all flows pass, this would mean the old front-end (step_streamline etc.) could be deprecated.

[fpjentzsch]

Summary of the current status of various example model flows:

Model	Current Status
VGG10	AssertionError: node_MultiThreshold_1: Signed output requires actval < 0
MNv1	Bad node spec for node. Name: QuantAvgPool2d_0 OpType: QuantAvgPool2d
Cybsec	Bad node spec for node. Name: BipolarQuant_0 OpType: BipolarQuant
cnv/tfc 1bit	Bad node spec for node. Name: BipolarQuant_0 OpType: BipolarQuant
cnv/tfc 2bit	Unsupported nodes remain in between FINN operators
kws	Unsupported nodes remain in between FINN operators
GTSRB	Bad node spec for node. Name: Unsqueeze_55 OpType: Unsqueeze
Transformer (benchmark)	dequant_softmax is unexpectedly None
Transformer (others)	Exception: Could not find a suitable int datatype for 8.711228073946339e+40