Fix type notation of merges in BPE Python binding

[Coqueue]

Hi, this change is to correct type notations of merges in Python binding of BPE as the counterpart in Rust implementation expects a file name string or Vec<(String, String)> instead of a mapping from a tuple of integers to another.

This is a second PR with the same change as PR#1685, which got closed with replies to questions from an reviewer that got ignored. I'll include the reasoning for this changes below.

Taking the Python SentencePieceBPETokenizer class as an example, the merges annotated as Optional[Union[str, Dict[Tuple[int, int], Tuple[int, int]]]] is only used to initialize the BPE class:

if vocab is not None and merges is not None:
        tokenizer = Tokenizer(BPE(vocab, merges, dropout=dropout, unk_token=unk_token, fuse_unk=fuse_unk))

where the BPE class is generated from the Rust implementation mentioned in the original post.

BPE class also contains below comment:

class BPE(Model):
    """
    An implementation of the BPE (Byte-Pair Encoding) algorithm

    Args:
        vocab (:obj:`Dict[str, int]`, `optional`):
            A dictionary of string keys and their ids :obj:`{"am": 0,...}`

        merges (:obj:`List[Tuple[str, str]]`, `optional`):
            A list of pairs of tokens (:obj:`Tuple[str, str]`) :obj:`[("a", "b"),...]`

This comment also gives a different type annotation for merges.

I encountered an issue initializing SentencePieceBPETokenizer with the given type annotation, but succeeded following the type annotation given in the comment of BPE class.

The type annotation of merges being Dict[Tuple[int, int], Tuple[int, int]]]] should actually be for MergesMap in Rust, as an alias of HashMap<Pair, (u32, u32)>, which is built internally in Rust from merges.

[ArthurZucker]

Sure! These files (in implementations) are kinda references, but are also not super maintained! Thanks 🤗