fix #144 for defaulting case, e.g. SciBERT without tokenizer config on HuggingFace

[kermitt2]

Inferring case from model name if available. This avoids BERT models on huggingFace not having a tokenizer config to default in do_lower_case=True even when it is a "cased" model.

[kermitt2]

Note that now when the DeLFT model is saved, the transformer tokenizer is also correctly saved.

For example with the problematic allenai/scibert_scivocab_cased, we have in the saved model transformer-tokenizer/tokenizer_config.json: "do_lower_case": false, as expected.

[lfoppiano]

Looks good. I just added a small change to lowercase the name, in case it's written with an uppercase letter - unlikely indeed

[kermitt2]

For reference, even with a model not really sensitive to case, this improves the result quite a lot with 10-folds:

python3 delft/applications/grobidTagger.py citation train_eval  --architecture BERT  --transformer allenai/scibert_scivocab_cased --input data/sequenceLabelling/grobid/citation/citation-231022.train --fold-count 10            

BEFORE CASE FIX             
            
Average over 10 folds
all (micro avg.)     0.9224    0.9479    0.9350 

AFTER CASE FIX 

Average over 10 folds
all (micro avg.)     0.9307    0.9526    0.9415

[lfoppiano]

Test with the average of 5 train and 5 evaluations with an holdout dataset:

Superconductors:

	Scibert	Scibert fixed	mattpu-scibert-initcheckpoint	mattpu-scibert-initcheckpoint fixed
<class>	72.74	71.08	71.37	71.65
<material>	78.84	80.04	78.82	80.17
<me_method>	66.75	65.88	65.42	65.73
<pressure>	42.04	47.15	43.66	44.62
<tc>	79.02	78.96	78.52	78.55
<tcValue>	78.74	78.10	78.46	78.63
All (micro avg.)	76.30	76.69	75.89	76.70

Quantities:

	Scibert	Scibert fixed	mattpu-scibert-initcheckpoint	mattpu-scibert-initcheckpoint fixed
<unitLeft>	93.09	91.80	92.14	92.66
<unitRight>	22.99	32.71	28.02	32.77
<valueAtomic>	81.89	86.55	86.13	87.06
<valueBase>	96.50	94.13	91.66	92.96
<valueLeast>	80.70	75.01	75.60	74.99
<valueList>	71.36	57.90	54.48	51.86
<valueMost>	78.61	72.41	73.59	72.51
<valueRange>	100.00	95.52	94.40	94.12
All (micro avg.)	84.82	85.15	85.09	85.53