[MO] relax condition for casting into FP16

[pavel-esir]

Root cause and solution 1: In some models -inf values are present in the original FW file, and these models after conversion into FP16 work correctly. But in #6590 strong condition for limit check was introduced. Need to relax limit check condition in ChangeOutputTypeAttributes.py.

Ticket: CVS-65735

Root cause and solution 2: In some models patterns input_1/Maximum(input_2, eps) and input_1/Add(input_2, eps) are used to prevent division to zero. But usually in FP32 networks eps is such small (e.g. 1e-9, 1e-12, ...) that after casting to FP16 it's collapsed to zero. This can lead to division to zero if input_2 is also zero. To prevent that we change eps to FP16 smallest normal value in such patterns in transformation DivisionToZeroFP16Resolver.py.
Moved solution 2 into nGraph and created separate PR #8676

Ticket: xxx-66045

Code:

• Comments
• Code style (PEP8)
• Transformation generates reshape-able IR
• Transformation preserves original framework node names

Validation:

• Unit tests
• Framework operation tests" n/a
• Transformation tests: n/a
• Model Optimizer IR Reader check: on Dien there are some problems but they are caused not by my changes

Documentation:

• Supported frameworks operations list: n/a
• Guide on how to convert the public model: n/a
• User guide update: n/a

[pavel-esir]

Communicated personally with @mvafin. He asked to return back unit-tests when casted to infinity and also add warning when casted into zero. Both requests addressed.

@rkazants please review

[rkazants]

@evolosen, please review it as well. Thanks!

[pavel-esir]

After @mvafin changes #7588 importance of transformation ChangeOutputTypeAttributes.py has been lowered.

Added also solution for division to zero in FP16 models. @sadolini @sevhabert @mvafin @rkazants please take a look one again to the latest changes.

[rkazants]

We discussed to proceed with different solution, namely, to fuse the sub-graph into MVN operation with original epsilon value.

[pavel-esir]

We discussed to proceed with different solution, namely, to fuse the sub-graph into MVN operation with original epsilon value.

As we agreed on Friday (19 Nov) moved small eps modifying transformation in specific patterns into nGraph with separate PR. But the current PR is still needed.

[pavel-esir]

Closing since old mechanism to convert to FP16 will not be supported in MO.