Fix quantizer propagation algo

[asenina]

Changes

Now before letting the quantizer pass through a branching node, we check that other branches coming out of the node should also be quantized. Otherwise the quantizer should not be moved.

Reason for changes

This PR fix bug(see tickets) and aligns the placement of quantizers in POT and NNCF.

Related tickets

#60615

Tests

Tests were added in test_quantizer_propagation_solver.py

[asenina]

Jenkins please retry a build

[ljaljushkin]

Jenkins please retry a build

[asenina]

Jenkins please retry a build

[wonjuleee]

test this please

[openvino-nncf-ci]

Can one of the admins verify this patch?

[vshampor]

The newly added test_intermediate_output before and after the fix:

Before:

After:

Note that the intermediate output is no longer quantized since the default behaviour is to have "quantize_outputs": False, while still having activations of both convs quantized.

[vshampor]

Hangs on test_quantize_network[w_sym_t_a_sym_t-mask_rcnn], probably due to excessive complexity of the algo implementation. Will try to optimize.

[vshampor]

TF's mask-rcnn quantization keeps hanging. The issue is not with the quantization algorithm - hang occurs in applying the transformation layout to the model.

The transformation layout before and after the PR has the same amount of quantizers, but some of these have changed positions so as not to quantize outputs (since "quantize_outputs" is false by default). Here's the diff of the transformation_layout.transformations list (left is before, right is after the PR):

259,263c259,263
< 0 TargetType.AFTER_LAYER rpn-box 3 0
< 0 TargetType.AFTER_LAYER rpn-box 4 0
< 0 TargetType.AFTER_LAYER rpn-box 0 0
< 0 TargetType.AFTER_LAYER rpn-box 1 0
< 0 TargetType.AFTER_LAYER rpn-box 2 0
---
> 0 TargetType.BEFORE_LAYER tf.reshape_7 0 0
> 0 TargetType.BEFORE_LAYER tf.reshape_9 0 0
> 0 TargetType.BEFORE_LAYER tf.reshape_1 0 0
> 0 TargetType.BEFORE_LAYER tf.reshape_3 0 0
> 0 TargetType.BEFORE_LAYER tf.reshape_5 0 0
310c310
< 0 TargetType.AFTER_LAYER box-predict 0 0
---
> 0 TargetType.BEFORE_LAYER tf.reshape_35 0 0

[vshampor]

Launched following builds: E2E PT build: 359 - no significant changes, UNet INT8 failure known E2E TF build: 366 - no significant changes, failures in the OV column are CI-related E2E ONNX build: 356 - no visible degradations "conformance" manual build: 58 - no visible degradations