Clean up public header for Python/Julia bindings

[ZedongPeng]

Description

interface.h currently includes many internal structs and functions, which makes it hard to build clean bindings and causes unnecessary CUDA dependencies.

We should separate public and private headers:

• Keep only the C API declarations in interface.h
• Separate internal/external structs and functions

This will make the public API cleaner and easier to maintain.

[ZedongPeng]

Closed via #26