fix flert hidden context breaks reduced vocab

[helpmefindaname]

closes #3368

[alanakbik]

Minimal script to reproduce the error:

import flair
from flair.data import Corpus
from flair.datasets import WNUT_17
from flair.embeddings import TransformerWordEmbeddings
from flair.models import TokenClassifier
from flair.trainers import ModelTrainer

flair.device = "cpu"
flair.set_seed(123)

# error thrown for [xlm-roberta-base, xlm-roberta-large, roberta-base], but not for [bert-base-uncased]
transformer_model = "roberta-base"

corpus: Corpus = WNUT_17().downsample(0.01)

label_dictionary = corpus.make_label_dictionary("ner")

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

tagger = TokenClassifier(embeddings=embeddings, label_dictionary=label_dictionary, label_type="ner")

trainer = ModelTrainer(tagger, corpus)

trainer.fine_tune(
    "./results/",
    max_epochs=1,
    reduce_transformer_vocab=True,
)

(I could reproduce the error for roberta-based models, but not bert-based models.)

[alanakbik]

Thanks @helpmefindaname for fixing this!