starcoder : add support for starchat special tokens

[marella]

cc @TheBloke

[TheBloke]

Nice!

[marella]

I noticed that the gpt_tokenize() function doesn't handle special tokens properly due to two issues:

1. It creates regex without escaping special characters in tokens, so it doesn't work for tokens like <|end|> which contains the special regex character |
2. It combines the special token pattern with the normal word pattern which greedily splits ' <|end|>' into ' <|', 'end', '|>'

So I updated gpt_tokenize() to handle this. For 1, I simply escaped the special characters. For 2, I first split the text by special tokens and then split the substrings in-between special tokens into words using the normal word pattern.