[Feature Request] [Training] Add Scan and other control flow operators

Nobody has claimed this yet.

### Describe the feature request

I want training support for mainly Scan, and possibly other control flow (If and Loop). Because Scan hold a subgraph to compute, we cannot use existing way to implement gradient for ops described in [here](https://github.com/microsoft/onnxruntime/wiki/Per-Op-Gradient-Building).
Implementing gradient support for Scan enables on-device training for RNN-based model (like RWKV, or Mamba) without custom kernel, so I think it will be great for less popular devices/accelerators (if ORT training support varieties of EP).

I think GetExternalGradient does gradient calculation by using custom ops and is considered as one of the solutions, but still documentations are not enough. (Sorry if it's wrong)

### Describe scenario use case

As mentioned above, on-device training without custom kernel, or on-device training with complex operations that cannot be expressed with CumSum/CumProd and other operators.

Start by reading the linked Per-Op Gradient Building wiki page and investigating GetExternalGradient, which the issue identifies as a possible approach. Clarify the intended training support for Scan and whether If and Loop are also in scope; done should include a defined implementation path and documented behavior for the requested control-flow operators.

Written by the indexing model from the issue text.