Pre-allocate space for Executorch inputs and outputs (#1385)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/1385

Allow use of pre-allocated input and output buffers for Executorch programs.
This requires no changes for inputs from python, but outputs from python need to
be pre-allocated.

This feature is always-on for now, and just catches errors if the outputs were already
pre-allocated. Doesn't seem like there's a need to add a new flag to the API.

Differential Revision: D51868693

fbshipit-source-id: d8d7a0c96aa4b24e07c36baaa0508b6546ca0f53