Extend to multiple foveation levels

Background: General frameworks for spatially adaptive generation benefit from support for varying degrees of resolution sparsification.

Question / Future Work: The current framework uses a spatial downsampling factor of $d=2$ (a $4\times 4$ computational gain in the periphery). Future work could explore extending the framework to support multiple levels of foveation, necessitating multi-level adaptation of phase-aligned Rotary Positional Embeddings (RoPE).