Implement category-based CER/WER

Create a new command atr-ner-eval cer --label label.bio --pred pred.bio to evaluate CER / WER for each category

Steps:

• for each ground truth/prediction pair
• for each category
• filter text by category
• compute CER/WER using the editdistance package

Expected output: a markdown table with CER/WER by category (you can use prettyTable)

| Category | CER (%) | WER (%) | N entities |
|----------|:-------:|:-------:|-----------:|
| person   |  5.61   |  24.15  |      1501  |
| date     |  3.50   |  19.24  |        62  |
| location |  1.35   |  12.40  |       101  |
| other    |  4.65   |  21.36  |     15048  |