SPARKNLP-921: Bug Fix for BPE and RobertaForQA

[DevinTDHa]

Description

This PR introduces several fixes to produce expected behavior of RobertaForQuestionAnswering models.

Resolves #14005.

It includes the following changes:

1. BPE Tokenizer includes a flag whether or not to always prepend a space before words (previous behavior for embeddings)
2. BPE Tokenizer correctly converts and tokenizes non-latin other special character words
3. RobertaForQA produces the same logits and indexes as the transformers implementation
   • Note: Spark NLP uses 1-based indexing. So to retrieve the answer strings, you have to use context.slice(start + 1, end + 1)

Motivation and Context

Users should expect the same results as their custom exported models

How Has This Been Tested?

New tests added, existing tests are passing

Screenshots (if appropriate):

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• Code improvements with no or little impact
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

• My code follows the code style of this project.
• My change requires a change to the documentation.
• I have updated the documentation accordingly.
• I have read the CONTRIBUTING page.
• I have added tests to cover my changes.
• All new and existing tests passed.