FWIW while two out of three filters that EXT_meshopt_compression provides are designed for quantized data (normals & quaternions), the exponential filter is designed for float32 data and can be used without KHR_mesh_quantization; gltfpack will do this when asked (using -vpf or -vtf flags for positions or texture coordinates respectively). Only noting this because it looks like this PR assumes filters must be applied to quantized accessors.

The value of using exponential filter is that it is transparent to the application (as it's just a transform that the rest of glTF document doesn't observe) but it allows to smoothly adjust compression ratio with some precision loss (controlled via bits & mode). The compression ratio (and VRAM savings!) are still the best with quantized data, so it's not a full replacement, but a useful alternative.

@zeux Really helpful, thank you! I see I have some incorrect assumptions in this PR — I may need to rethink the API currently exposed for configuring EXT_meshopt_compression.