apache/tvm-ffi ? reverse-engineered prompt

Build me a small open source library for machine learning tools that makes it easy to call functions and pass tensors between Python, C++, and Rust without copying data around.

I want a clean, stable C level interface, plus friendly wrappers for the main languages, so someone can ship one package that works across different ML frameworks and runtimes. It should feel practical for things like kernel libraries, custom operators, and code generated by tools. Please include a few simple examples, basic tests, and clear docs for getting started, building from source, and using it from each language. If you need to check the latest docs or best practices for packaging and interoperability, look them up online first.

Make it polished enough that another developer can clone it, build it, and understand how to export functions and classes, share tensors safely, and use it in a real project.