BC3 alpha rounding behavior

[neptuwunium]

Following the discussion n the iOrange/bcdec#12: BC3 the alpha channel is interpolated as a floating point channel in modern implementations.

Seemingly, at some point BC3 split into two (likely with the introduction of BC3 into DXGI vs the DXT5 format.)
One that uses smooth_alpha_block to interpolate alpha, and a modern variant that uses bc4_block.

The resulting call would look like

  if modern_spec {
    bc4_block(
        compressed_block,
        &mut decompressed_block[3..],
        destination_pitch,
        4,
        false,
    );
  } else {
    smooth_alpha_block(
        compressed_block,
        &mut decompressed_block[3..],
        destination_pitch,
        4,
    );
  }

Additionally, smooth_alpha_block is implemented inaccurately as-is, see related commit: iOrange/bcdec@963c5e5#diff-a7d40b7cf11635d27584cfdad0351fef5fce1c22fe1ad20060b078ecbd175a1aL209

[ScanMountGoat]

This is related to the discussion for #17. There is some error allowed by the specifications I've found. The GPU and software decoders I've tested often differ by 1.0/ 255 for BCN formats other than BC6 and BC7. I'm not against improving the accuracy relative to a floating point implementation as long as the performance is similar and the test suite still works.