Add a note about padded vocab size during training

Per our Slack convo — for newcomers to BLOOM it's unclear why the vocab_size in the config.json is 250880 while the model card says 250680. This helps clarify that the effective vocab size is 250680 but the instantiated matrix will be of shape 250880.

README.md

@@ -193,6 +193,8 @@ The BLOOM tokenizer ([link](https://huggingface.co/bigscience/tokenizer)) is a l
 
 - A vocabulary size of 250,680
 
+The vocabulary size was padded to 250,880 for practical purposes during training, but the effective model vocabulary size is 250,680.
+
 It was trained on a subset of a preliminary version of the corpus using alpha-weighting per language.    
     
 </details>