Token indices sequence length is longer than the specified maximum sequence length for this model

[LysandreJik]

As of now the tokenizers output a specific warning when an encoded sequence is longer than the maximum specified sequence length, which is model-specific:

Token indices sequence length is longer than the specified maximum sequence length for this model (X > 1024). Running this sequence through the model will result in indexing errors

It is currently in the convert_tokens_to_ids and this leads to two issues:

• using encode or encode_plus methods with a max_length specified will still output that warning as the convert_tokens_to_ids method is used before the truncation is done. (cf token indices sequence length is longer than the specified maximum sequence length  #1791)
• since prepare_for_model was introduced, I personally feel that all modifications related to the model should happen in that method and not in tokenize or convert_tokens_to_ids.

This PR aims to slightly change the behavior so that both aforementioned issues may be solved by putting the warning in the prepare_for_model method if no max_length is specified.

[thomwolf]

LGTM, cleaner indeed!