Squirrel across tflite tensor type in `iree.reflection`

We need the tflite canonical type for each I/O tensor. They're not just normal primitives but also `kTfLiteComplex64` and such. For now float32 is fine, but quantization #3972 will require the actual type.

Start by locating the iree.reflection handling for TFLite I/O tensors and trace how their types are currently represented. Review the relationship to quantization issue #3972 and the TFLite types mentioned here. Done means each I/O tensor exposes its canonical type, including types such as kTfLiteComplex64 rather than assuming float32.

Written by the indexing model from the issue text.