Introducing TF v2 wrapper, ModelSignatureManager, and easy import of TF v2 models

[maziyarpanahi]

This PR introduces the following features:

• Updates existing loadSavedModel function to import TensorFlow SavedModel to be compatible with TF v2
• Updates existing read and write TF v1 graph (freeze) to be compatible with TF v2 (currently crashes for none existing targets in TF v2 graph)
• Implements ModelSignature to support different model providers such as TF Hub, HuggingFace, etc. (if they are for raw embedding with last_hidden_state or pooler_output) for SavedModel in TF v2 (currently only supports TF v1 with hard-coded placeholders for inputs and outputs)

Annotators adapted to these changes to support TF v2 SavedModel:

• BertEmbeddings
• BertSentenceEmbeddings
• AlbertEmbeddings
• XlnetEmbeddings
• DistilBertEmbeddings

NOTE: All the previous models exported from TF v1 or uploaded as pretrained are still compatible. This PR will extend the support to TF v2 SavedModel especially the use of loadSavedModel in Python/Scala to easily import models from TF Hub and HuggingFace for the listed annotators.

NOTE2: There are other branches using TF v2 models which rely on this branch. (it was the only way to have support for TF v2 SavedModel for development)

########
This PR introduces a new annotator for RoBertaEmbeddings based on RoBERTa models (for instance roberta-base).

Import the following models:

• roberta-base
• roberta-large
• distilroberta-base

########
This PR introduces a new annotator for DistilBertEmbeddings based on DistilBERT models (for instance distilbert-base-cased).

Benchmarks against bert_base_cased have shown better speed and very close accuracy in NER (faster convergence).

Import the following models:

• distilbert-base-cased
• distilbert-base-uncased
• distilbert-base-multilingual-cased

########

Other PRs

• BPE, input IndexToken, Output Array[TokenPiece] #2867 first implementation of BPE tokenizer for roBERTa, GPT2, and XLM
• Introducing DistilBertEmbeddings annotator based on DistilBERT models  #2871 DistilBertEmbeddings annotator
• Introducing RoBertaEmbeddings annotator based on RoBERTa models #2902 RoBertaEmbeddings annotator
• Added fix for ending zero index in single char initial token in positional mask algorithm #2874