[FEA] Support different models in sequence classifier

[efajardo-nv]

Is your feature request related to a problem? Please describe.
Sequence classifier only uses bert-base-uncased base model and corresponding vocabulary and tokenizer for fine tuning.

Describe the solution you'd like
Allow use of different base models and way to pass corresponding vocabulary hash file for cudf subword tokenizer.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed.