Fix training sample batch2

[exfalsoquodlibet]

Summary

Hack-y code to fix the (incremental) training set after the sentences were accidentally segmented for step 2 when we annotated for the extra 8 categories: GPE, ORG PN, PERSON PN, POSTCODE, EMAIL, PHONE N, DATE, MONEY £

This requires a hack to ensure we could merge these annotations with the original set annotated for FORM.

This code addresses this. And also de-duplicated the set.

Checklists

This pull/merge request meets the following requirements:

• [x] code runs
• [x] developments are ethical and secure
• [x] you have made proportionate checks that the code works correctly
• [ ] test suite passes
• [ ] developments adhere to AQA plan (see docs/aqa/aqa_plan.md)
• [ ] data log updated (see docs/aqa/data_log.md), if necessary
• [ ] assumptions, and caveats log updated (see docs/aqa/assumptions_caveats.md), if necessary
• [ ] minimum usable documentation written in the docs folder

Comments have been added below around the incomplete checks.