Fixed extracting TF models with string constants

[lazarevevgeny]

Root cause analysis: After merge of #3573 the MO starts to read operations located in the sub-graphs of the main model (While operation body). There is an operation inside containing string constant value. MO fails to read this constant value because it tries to convert it to numpy string and fails.

Solution: Do not try to convert the value to numpy string and use it as is. Similar fix was done for the issue before but this particular line of code was not updated.

Ticket: 47405

Code:

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

Validation:

• Unit tests
• Framework operation tests
• Transformation tests: N/A
• Model Optimizer IR Reader check: done manually

Documentation:

• Supported frameworks operations list: N/A no new op enabled
• Guide on how to convert the public model: the instruction exists already
• User guide update: N/A the instruction exists already

[pavel-esir]

Looks good to me. Looks like some tests failed because of testing infrastructure issues not because of changes in the PR