Original code didn't produce the expected results when multiple skinned mesh renderers were used. Only the first mesh was transformed correctly, and the rest seemed to follow the first one's transformation.
Changes proposed in this fork are related to a jobified transformation step right before the vertices are passed to the ComputeBuffers, and also on initialization.
Core functionality remained untouched
Considerable performance improvements can be achieved by moving the transformation to the Compute Shader, right before the positions are hashed.
With respect, George Adamopoulos