allow finetuning of word-embeddings

[helpmefindaname]

This PR reworks the WordEmbeddings to add a few features:

• embeddings are stored as a torch Embedding instead of a gensim keyedvector. That way it will never come to version issues, if gensim doesn't ensure backwards compatibility (happened to me a few times over the last 2 years)
• embeddings can be put on gpu, if wanted. For that, WordEmbeddings('<name>', force_cpu=False) can be used
• embeddings can be fine-tuned. For that, FordEmbeddings('<name>', fine_tune=True) can be used.
• the lru_cache will store indices per word instead of vector lookups. This is because the lookups shouldn't take long if calculating all indices at once (especially if using gpu). However, the regex operations and multiple lookups will remain cached.
• lru_cache is increased, as it takes (<embedding_length> - 1)* 4 bytes less per lookup.
• allows usage of "stable embeddings" → https://arxiv.org/abs/2110.02861 which seem to "reduce gradient variance that comes from the highly non-uniform distribution of input tokens"

[alanakbik]

@helpmefindaname thanks a lot for adding this, really cool features!

I ran some experiments with the vanilla Flair configuration on CoNLL-03 (glove+flair embeddings) with three settings: no-finetune (i.e. same as before), finetune-normal (with finetuning) and finetune-stable (with finetuning and stable=True). Each configuration ran 5 times, here are averages and standard deviation:

Approach	Test F1
no-finetune	93.01 +- 0.12
finetune-normal	93.03 +- 0.08
finetune-stable	92.76 +- 0.04

[alanakbik]

@helpmefindaname we just realized that gensim is pinned to 3.8. max in Flair, which does not work with Python 3.9.

So I was thinking of upgrading to gensim version 4, but there's lots of changes in the syntax, so I guess that would mean that 3 no longer works. Do you think requiring gensim > 4 makes sense in Flair?