Support for 4bit or 8bit tensor

Although it may be out of scope, it would be nice to have an example of computing 4bit and 8bit tensors, to save memory bandwidth.

No files, tests, or entry points are named. First clarify the intended 4-bit and 8-bit tensor representation and the example's expected scope, then identify the relevant tensor and GPU-computation entry points; done should include a working example that demonstrates the requested reduced memory bandwidth.

Written by the indexing model from the issue text.