[Fix] Inferring a default sequence length during transformers export

[dbogunowicz]

This PR: #1826 has added new functionality so that when no sequence_length is being specified, we are defaulting to a max sequence length from the config. Unfortunately, as discussed in the PR comments of 1826, indeed - not only an HF config is not guaranteed to have max_position_embeddings attribute, but this information can be present under a different key value.
E.g
For the model TinyLlama-1.1B-Chat-v0.3 we are looking at config.max_position_embeddings
For the model zoo:mpt-7b-mpt_chat_mpt_pretrain-pruned80_quantized we are looking at config.max_seq_len

This PR:

1. Adds max_seq_len to the set of config attributes that may be potentially used to infer the default sequence_length
2. Adds the behaviour that raises a ValueError if we are not able to infer the default sequence_length (the user should specify it manually then)

[derekk-nm]

Could a config have both attributes? If so, is max_position_embeddings preferred over max_seq_len, as coded here?

[dbogunowicz]

Good catch. Regarding both keys being simultaneously present- honestly do not know, I would need to go through more example configs. But preferring max_position_embeddings or max_seq_len sounds reasonable to me.