catena-gpu naive matmul

[mstn]

Summary

Naive matmul for catena-gpu. See doc for details.

Change to catena-lang runtime

We want to use the same runtime as catena-lang because it is currently used in other parts of the stack. We needed to change it.

It was primarily an API extension, but I did touch the implementation of an existing method.

Added APIs:

• Public GeneratedFunction ABI metadata.
• Runtime::dialect().
• Runtime::load_generated_source(source, functions).
• Existing public method signatures remain unchanged.

I added new method Runtime::load_generated_source(source, functions)

Existing load methods accept Hex/Catena source and always run the complete catena-lang pipeline:

Runtime::load(...)
Runtime::load_sources(...)

That cannot load catena-gpu output. By the time catena-gpu calls the runtime, it already has generated CUDA/HIP and ABI signatures.

Runtime::load_generated_source(source, functions) adds some code duplication, but it is the safest path that does not require to change/break existing code.

Questions

Hardware-dependent constraints include:

• each block axis within maxThreadsDim;
• block volume within maxThreadsPerBlock;
• each grid axis within maxGridSize;
• kernel-specific register and shared-memory resource limits.

should we provide safety or is it fine if it crashes?