GH-2015: flert

[alanakbik]

This PR adds FLERT as presented in our recent paper (closes #2015).

A number of changes are made:

1. Sentence objects now have next_sentence() and previous_sentence() methods that are set automatically if loaded through ColumnCorpus. This is a pointer system to navigate through sentences in a corpus:

# load corpus
corpus = MIT_MOVIE_NER_SIMPLE(in_memory=False)

# get a sentence
sentence = corpus.test[123]
print(sentence)
# get the previous sentence
print(sentence.previous_sentence())
# get the sentence after that
print(sentence.next_sentence())
# get the sentence after the next sentence
print(sentence.next_sentence().next_sentence())

This allows dynamic computation of contexts in the embedding classes.

2. Sentence objects now have the is_document_boundary field which is set through the ColumnCorpus. In some datasets, there are sentences like "-DOCSTART-" that just indicate document boundaries. This is now recorded as a boolean in the object.

3. TransformerWordEmbeddings refactored for dynamic context, robustness to long sentences and readability. The names of some constructor arguments have changed for clarity: pooling_operation is now subtoken_pooling (to make clear that we pool subtokens), use_scalar_mean is now layer_mean (we only do a simple layer mean) and use_context can now optionally take an integer to indicate the length of the context. Default arguments are also changed.

For instance, to create embeddings with a document-level context of 64 subtokens, init like this:

embeddings = TransformerWordEmbeddings(
    model='bert-base-uncased',
    layers="-1",
    subtoken_pooling="first",
    fine_tune=True,
    use_context=64,
)

From my testing, it also seems that the new implementation is a bit faster.

4. You can train FLERT like this:

import torch

from flair.data import Sentence
from flair.datasets import CONLL_03, WNUT_17
from flair.embeddings import TransformerWordEmbeddings, DocumentPoolEmbeddings, WordEmbeddings
from flair.models import SequenceTagger
from flair.trainers import ModelTrainer


corpus = CONLL_03()

use_context = 64
hf_model = 'xlm-roberta-large'

embeddings = TransformerWordEmbeddings(
    model=hf_model,
    layers="-1",
    subtoken_pooling="first",
    fine_tune=True,
    use_context=use_context,
)

tag_dictionary = corpus.make_tag_dictionary('ner')

# init bare-bones tagger (no reprojection, LSTM or CRF)
tagger: SequenceTagger = SequenceTagger(
    hidden_size=256,
    embeddings=embeddings,
    tag_dictionary=tag_dictionary,
    tag_type='ner',
    use_crf=False,
    use_rnn=False,
    reproject_embeddings=False,
)

# train with XLM parameters (AdamW, 20 epochs, small LR)
trainer = ModelTrainer(tagger, corpus, optimizer=torch.optim.AdamW)
from torch.optim.lr_scheduler import OneCycleLR

context_string = '+context' if use_context else ''

trainer.train(f"resources/flert",
              learning_rate=5.0e-6,
              mini_batch_size=4,
              mini_batch_chunk_size=1,
              max_epochs=20,
              scheduler=OneCycleLR,
              embeddings_storage_mode='none',
              weight_decay=0.,
              )