SDL_GPU: Prevent allocating/copying uniform buffers when possible

[flibitijibibo]

sdlgpu.c's update_uniform_buffer function stands out in profiles of FNA3D trace data. For OpenGL we were able to reduce glUniform's impact on CPU performance by memcmp'ing the buffer first, and only pushing when a change actually happened...

https://github.com/icculus/mojoshader/blob/main/mojoshader_opengl.c#L2813

... but sdlgpu.c currently allocs a buffer for every constant buffer push:

https://github.com/icculus/mojoshader/blob/main/mojoshader_sdlgpu.c#L195

We should try to remove this middle buffer and shadow buffer state, this can improve performance quite a bit!

[flibitijibibo]

Latest revision removes most of the malloc calls and a lot of the other goo that was in update_uniform_buffer. The remaining issue is the memcpys, which mostly comes down to figuring out the scope of the uniform buffer state and shadowing state accordingly.