Store compressed data relative to the bind pose translation

[nfrechette]

In the spirit of bind pose stripping, we can leverage the fact that the output pose buffer contains the bind pose to store our compressed data relative to it. See here for the general idea.

Storing the compressed pose to be a full additive on top of the bind pose is expensive to reconstruct at runtime as it requires quaternion multiplication. Instead, we could store relative to the bind translation only by removing it from our raw data and adding it back after decompression. This is much faster and can be accomplished with 2 uOps: a vector addition and a memory load.

This would reduce the range of translation values to be centered around 0.0 which should help improve precision and reduce the memory footprint.

However, much like bind pose stripping, this cannot be supported for bones that are individually decompressed. In fact, if any bone is individually decompressed, we cannot support this. We could exclude the root but not arbitrary bones as easily.