Preserve disconnected Constant->Result subgraphs in TensorFlow models

[evkotov]

Details:

After CVS-136813, TensorFlow Frontend stopped removing Constant->Result subgraphs that represent training artifacts (e.g., Adam optimizer parameters like beta_1, beta_2). However, the test ModelWithConstResultSubgraphs was failing because:

1. ConstToResultRemover was incorrectly removing all Constant->Result subgraphs, not just UnsupportedConstant->Result
2. The reference model in the test didn't include the disconnected constants

Fixed ConstToResultRemover transformation
src/frontends/tensorflow_common/src/helper_transforms/const_to_result_remover.cpp

Removed logic that deleted all Constant->Result subgraphs. Now only removes UnsupportedConstant->Result subgraphs as intended. This preserves legitimate training artifacts from TensorFlow models.

// Before: removed both UnsupportedConstant and Constant
if (unsupported_const || const_node) {
    results_to_remove.push_back(result);
}

// After: only remove UnsupportedConstant
if (unsupported_const) {
    results_to_remove.push_back(result);
}

Updated reference model
src/frontends/tensorflow/tests/convert_tricky_models.cpp

Added disconnected Adam/beta_1 and Adam/beta_2 constants to the reference model to match the actual TensorFlow model structure.

Tickets:

• 140357

[rkazants]

we have a model that contains multiple isolated Constants (tens of them) that impact compilation times of the plugin.
This was a reason to delete such isolated constants.

[rkazants]

we have a model that contains multiple isolated Constants (tens of them) that impact compilation times of the plugin.
This was a reason to delete such isolated constants.