[NestedTensor] RelaxedUnspecConstraint failures due to mark_dynamic in NT constructor about pytorch

[davidberard98]

Ideas:

1. Change model code so that all tensors with the given dimension are marked as dynamic. Note: I don't this isn't as simple as just mark_dynamic() on a bunch of tensors, because we need to make sure the mark_dynamic is applied right before the graph(s) that have issues. Instead, @YuqingJ found that we can wrap the other dense tensors in NestedTensors and use their values, which is effectively the same as forcing a graph break & marking dynamic.
2. Don't mark_dynamic: ... this means we need to fix NJT flatten/unflatten to work with non-dynamic values/offsets. Also, we may need to introduce an API to turn this on/off. Note that the API needs to be somewhat global; we need to control not only the initial construction of the NJT, but also the results of operations on NJTs (which also construct NJTs internally in their implementations in ops.py)
3. Relax RelaxedUnspecConstraint asserts or introduce a new API in dynamic shapes: conceptually, for non-export, it seems like mark_dynamic may be too strict. In the NJT (non-export) use case, mark_dynamic is more of a suggestion to reduce the amount of recompilation we do. However, it's not necessary; and practically, in most current NJT enablement work, it's probably true that most graphs dealing with similar jagged data will be recompiling anyway. For this use case, it seems like it would be better to not assert here. Would it be reasonable to add the assert only if we're exporting, or add a different API (e.g. mark_dynamic(..., enforce=False)) which has this behavior?