Fix bug with CoT early stopping and LLama2 tokenizer

[bmosaicml]

What does this PR do?

LLaMa2 tokenizer didn't work with early stopping because it inserted some extra token (28971) before \n\n. This caused the stop sequence to appear longer than it really was, which resulted in the stopping criteria looking back too far into the input and causing generation to stop as soon as it begins. With the new implementation we remove any superfluous tokens in the stopping criteria (i.e. those tokens that add nothing to the decoded result) I also reduced the amount of extra lookback we do to make it so we don't ever look back farther than we should and stop as soon as generation begins.

Testing: mpt-eval-NzfmSy

| Category   | Benchmark   | Subtask   |   Accuracy | Number few shot   | Model                    |
|:-----------|:------------|:----------|-----------:|:------------------|:-------------------------|
|            | gsm8k       |           |   0.141774 | 0-shot            | meta-llama/Llama-2-7b-hf |

Results replicate LLama2 paper (14.6 for 7b)

What issue(s) does this change relate to?

Before submitting

• Have you read the contributor guidelines?
• Is this change a documentation change or typo fix? If so, skip the rest of this checklist.
• Was this change discussed/approved in a GitHub issue first? It is much more likely to be merged if so.
• Did you update any related docs and document your change?
• Did you update any related tests and add any new tests related to your change? (see testing)
• Did you run the tests locally to make sure they pass?
• Did you run pre-commit on your change? (see the pre-commit section of prerequisites)

[dakinggg]

Can you please include some evidence this resolves the issue in the pr description?