[Feature]: AutoDeploy: discussion where/how/when to enforce strict export IR

🚀 The feature, motivation and pitch

Right now, we technically impose a strict requirement to keep our fx graph consistent with the export IR specifications. However, we already break this assumption in multiple places and hence should revisit this assumption

1. Quasi-dynamic shapes in prepare_metadata and using batch_info tensor to split sequences
2. In-place update with cuda_cached_causal_conv
3. Irregular function for gather_logits with undeclared dynamic shape

Let's revisit the assumption of enforcing a strict export IR during/after transforms!

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.