Simpler and faster encoding

I think the logic in encode can be simplified, making it easier to understand, while meaningfully improving performance.

The current function, in my understanding, should have runtime proportional to 3*m*n where m is the length of merges and n is the length of text.

def encode(self, text):
    text_bytes = text.encode("utf-8") 
    ids = list(text_bytes)
    while len(ids) >= 2: # should loop ~m times
        stats = get_stats(ids) # O(n)
        pair = min(stats, key=lambda p: self.merges.get(p, float("inf"))) # O(n)
        if pair not in self.merges:
            break
        idx = self.merges[pair]
        ids = merge(ids, pair, idx) # O(n)
    return ids

We can simplify the logic significantly by leveraging the existing merge function and just merging over every pair in merges. This should also have a runtime proportional to 1*m*n:

def faster_encode(self, text):
    text_bytes = text.encode("utf-8")
    ids = list(text_bytes)
    for pair, idx in self.merges.items(): # O(m)
        ids = merge(ids, pair, idx) # O(n)
    return ids

In testing with various vocab sizes and input text length, I'm seeing a ~2-3x speed improvement for faster_encode, and it seems like this logic can be extended to the other Tokenizer classes.