Fix K-bits range for meshopt octahedral filter

[lilleyse]

Looking at meshopt_encodeFilterOct as a reference, the bit range should be 1 <= K <= 16, not 4 <= K <= 16.

@zeux does this change look correct?

[zeux]

While 1 is the smallest possible K from the encoding perspective, unsure if the spec should state that as the bounds - a 1-bit signed normalized integer has two values, -1 and 0, so that's not super useful for normal encoding :) My recollection for why the spec stated K=4 is the minimum was that I felt that this is the lowest value assets could be reasonably encoded in (at a great quality loss!), but maybe this is too conservative. When normals are used for specular reflections, Suzanne from glTF-Sample-Assets looks broken on K=3:

K=4

K=3

... but BrainStem model is perhaps okay for some contexts with K=3:

... or K=2

A 2-bit signed normalized integer can represent values -1, 0, 1 in [-1, 1] range. So if we want to update the spec then minimum K should probably start with 2; K=1 is not useful.

[emackey]

@lilleyse What's the use case here, is it 3D Tiles with unlit and hence no normals are being used? What's the next step here?

[lexaknyazev]

It's a request to align the extension spec with the implementation. I think the next step is to update the PR to use 2 as the lower bound.