If the output has a symbolic dimension
  * Infer the shape if it is a loop state variable as we have the input value, and the shape from the subgraph output must match
  * Use a temporary MLValue for the first subgraph execution if it is a subgraph output with a symbolic dimension.
    * After the first execution make the overall output shape concrete and allocate the full output buffer.
    * Use slices of the full output buffer for all other subgraph executions to avoid copies.

Add unit test to validate.