Delay translating DCT tokens into coefficients until immediately before IDCT

This is generally around 12% faster than the prior method of creating a
linked list for each block as tokens are read, but can be anywhere from
8% to 28% faster depending on file and CPU.

git-svn-id: file:///var/local/repositories/ffmpeg/trunk@22190 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b