[NeoML] Reference DNN with composites fix

[favorart]

ChangeList:

1. CCompositeLayer - forwarded isReferenceDnn and isLearningEnabled flags to the internal dnn
2. CMultiheadAttentionLayer may have non-connected output, allow it in optimization
3. CMultichannelLookupLayer - apply SetUseFrameworkLearning( true ) for share the state
4. CDistributedInference - avoid assert calls on some empty results
5. Channelwise-layers may contain empty paramBLobs
6. ReferenceDnnFactory does not close an archive
7. Update reference dnn tests to check for composite layers