Truncating the prefix of a sequence rather than the suffix

# 🚀 Feature request

Hi, tokenizers get `truncation` as an argument. When set to `True` the tokenizer will truncate the suffix of a sequence so it does not surpass the specified `max_length`. I'd like to have a functionality that truncates the prefix of the sequence, so the model will see the suffix of the sequence.

## Motivation

In many applications (e.g. Dialog, and QA) the most important part of the sequence is the suffix (e.g. the question after the context, or the last response of the dialog).

## Your contribution

Perhaps I'll submit a PR, but it might take me some time as I'm close to some deadlines of mine :(

Start by tracing the tokenizer truncation argument through the Transformers tokenizer entry points and existing truncation tests. Determine how prefix truncation should interact with max_length and the existing suffix-truncation behavior. Done means a supported option preserves the sequence suffix and tests cover the requested dialog and question-after-context use cases.

Written by the indexing model from the issue text.