Adding CrowsPairs task for English and French

[oskarvanderwal]

I can successfully run the CrowS-Pairs (multilingual) tasks for the prompts we have written.

For python main.py --model gpt2 --device cpu --tasks crows_pairs_english I get:

gpt2 (), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
|       Task        |     Prompt      |Version|Metric|Value |   |Stderr|
|-------------------|-----------------|------:|------|-----:|---|-----:|
|crows_pairs_english|                1|      0|acc   |0.5092|±  |0.0122|
|crows_pairs_english|                2|      0|acc   |0.5069|±  |0.0122|
|crows_pairs_english|                3|      0|acc   |0.5069|±  |0.0122|
|crows_pairs_english|                4|      0|acc   |0.5194|±  |0.0122|
|crows_pairs_english|A_preference     |      0|acc   |0.4764|±  |0.0122|
|crows_pairs_english|A_stereotype_true|      0|acc   |0.4949|±  |0.0122|

There are no official implementations of the CrowS-Pairs benchmark that work for autoregressive models like GPT-2.
For another implementation of CrowS-Pairs (older version though), I get a bias score of 0.593501326259947 for GPT-2; So quite a bit higher---but that doesn't say too much since the operationalization of the measures is so different.

For python main.py --model gpt2 --device cpu --tasks crows_pairs_french I get:

|       Task       |       Prompt       |Version|Metric|Value |   |Stderr|
|------------------|--------------------|------:|------|-----:|---|-----:|
|crows_pairs_french|A_preference_fr     |      0|acc   |0.4997|±  |0.0122|
|crows_pairs_french|A_reality_check_fr  |      0|acc   |0.5134|±  |0.0122|
|crows_pairs_french|A_stereotype_true_fr|      0|acc   |0.5224|±  |0.0122|

[StellaAthena]

Great work, thanks for the PR