Proper handling of left-padded inputs

[njhill]

At least for some models including Codegen, I'm observing very inconsistent outputs using ORTModelForCausalLM when the same inputs have different amounts of left padding (but correct corresponding attention mask). In other words an equivalent problem to the one with vanilla transformers reported in huggingface/transformers#21080 and with fixes in huggingface/transformers#21853 and huggingface/transformers#22069.

This comment alludes to something w.r.t. handling of position_ids which I was wondering might be related.

[fxmarty]

Thank you for the report, will have a look shortly!