Recompose QLinearMatMul and remove Quantize-Dequantize pairs

[tungld]

Add two patterns found when doing static quantization for bert-base models using onnxruntime.

• 1st pattern: recompose DequantizeLinear - MatMul - QuantizeLinear back to QLinearMatmul
• 2nd pattern: remove QuantizeLinear-DequantizeLinear pairs

[tungld]

Remove the generated .so during make check-onnx-lit.

[AlexandreEichenberger]

LGTM, worthwhile opt, thanks

[jenkins-droid]

Jenkins Linux s390x Build #15146 [push] Recompose QLinearMatMul ... started at 01:12

[jenkins-droid]

Jenkins Linux amd64 Build #15141 [push] Recompose QLinearMatMul ... started at 00:12

[jenkins-droid]

Jenkins Linux ppc64le Build #14171 [push] Recompose QLinearMatMul ... started at 01:24

[jenkins-droid]

Jenkins Linux amd64 Build #15141 [push] Recompose QLinearMatMul ... failed after 24 min

[jenkins-droid]

Jenkins Linux s390x Build #15146 [push] Recompose QLinearMatMul ... passed after 1 hr 29 min

[jenkins-droid]

Jenkins Linux ppc64le Build #14171 [push] Recompose QLinearMatMul ... passed after 2 hr 0 min

[mgehre-amd]

The QuantizeDequantizePattern has broken out downstream use case.
It's wrong to fold DequantizeLinear(QuantizeLinear(x)) to identity because the rounding then doesn't happen anymore.
(and already ignoring that they might have different scales & zero-point cannot be right).

Example assuming scale = 1 and zero_point = 0:
DequantizeLinear(QuantizeLinear(3.5)) = DequantizeLinear(4) = 4.0

Can you please revert this?

[AlexandreEichenberger]

@tungld Can you look into @mgehre-amd comment and fix appropriately? Thanks

[tungld]

@mgehre-amd thanks for pointing it out! I created a PR to revert that pattern: #2952. Thanks!