Support for models larger than 2GB

[robertknight]

RTen is currently limited to models that are 2GB or less in size. This limitation is inherited from the FlatBuffers format which uses 32-bit offsets internally (see google/flatbuffers#7537). Since the only dtype that is currently supported for weights is f32, this means models are limited to ~500M parameters.

This was fine for the original use case of the engine, but limits its usefulness as a general ONNX runtime for modern models. When loading ONNX models directly (#141), this issue is solved in the spec by allowing models to reference external data files. For the .rten format, a different solution will be needed.