- do start position computation and 0xFF 0x00 escaping on the CPU using SIMD
- use `override` in the shader once naga supports that, and make the shader more flexible
- implement a faster huffman decoding strategy
  - play with loading 256 Bytes of the bitstream per thread into LDS, and, if needed, performing
    more than 1 shader pass (requires using `dispatch_workgroups_indirect`)