Fix error when decoding a token in the id gap (or out of range) in a tiktoken tokenizer

[dakinggg]

Previously, if you tried to decode an invalid token, the tokenizer would crash. For a trained model, this would never be an issue because it would not produce invalid tokens, but for an untrained model, it could randomly produce any token up to the embedding size. This PR has the tokenizer instead return empty string for these invalid tokens.

This behavior matches some Hugging Face fast tokenizers, but not slow tokenizers, which error on out of range indices. Even though the tiktoken wrapper is technically a slow tokenizer, the fast behavior seems better to avoid crashes on random models.

It also adds a utf-8 encoding to two file open calls. See mosaicml/composer#2824 for more details.

[sashaDoubov]

LGTM! this is very useful for debugging + thank you for digging into the utf-8 encoding issue.