SDL_GPU: Use separate command buffers for copy/render, shared transfer buffers

[MrBrixican]

• Separate command buffers for copy/render commands to allow for better throughput
• Now use shared transfer buffer when reasonable (<16MB) and single use transfer buffers when larger
• Lots of this is based on the SDL_GPU driver in FNA3D, and likely needs review
  • Aside from the TODO comments added here, I've made some notes in Refactor rendering to use SDL3 GPU #1
• Overall these changes allow for better throughput with synthetic benchmark performance on par with (actually a little better than) the current OpenGL implementation in main