Reduce overhead of CUDA actors

[fahim-hamid]

Summary of changes:

• Reduced CAF CUDA overhead in command submission, argument packing, lookup, allocation, and temporary command handling.
• Improved CUDA context lifetime, stream ordering, resource teardown, and module reload behavior.
• Batched synchronous output collection and removed unnecessary host/device copies where possible.
• Made CUDA memory ownership and callback handling safer.
• Updated SC26 sequence benchmarks to use consistent native/CAF semantics.
• Added CUDA 12/13 and Windows/MSVC compatibility fixes.
• Added automatic cubin, DLL, and executable output handling.
• Updated benchmark runners to support CMake builds and calculate averages/stddev.
• Added portable CMake targets for all SC26 benchmarks w/ instructions libcaf_cuda/sc26/benchmarks/README.md.
• Preserved the existing Makefile and legacy benchmark-runner workflows where practical, hopefully any personal scripts don't break, let me know if I can fix anything!