Add registry system to operators

Context

This MR refactors all operator implementations to use the Registrar system (similar to backend_cpu) instead of manual instantiation and type dispatch. Previously, each operator implementation contained manual if/else dispatch on DataType to call templated forward_ and backward_ functions.

With this refactor:

• Operators are now defined as type aliases of a generic OperatorImpl_cuda<Op,...> template.
• dtype-specific kernels are moved into dedicated *_CUDA_kernels.hpp files.
• Each kernel is registered once via the REGISTRAR macro for supported data types(It can also be specific to data format, specific attributes values ...).

It resolves #35 (moved).