feat(runner): replace CLI with Python JSON protocol

[tetsuo-cpp]

Summary

• replace the C++ CUDA runner with a cuda-python JSON stdin/stdout runner
• send inline base64 PTX and typed parameters from the Vast.ai test harness
• support structured errors and multiple output buffers
• remove remote nvcc compilation and the CMake runner target

Verification

• uv run pytest -q -m 'not gpu' (13 passed, 37 deselected)
• cmake --build build --target check-warpforth -j2 (117 passed)
• uv run ruff check gpu_test/
• uv run ruff format --check gpu_test/
• cmake --build build -j2

Closes#50