This removes one of the transposes from the fDCT, and avoids several zig-zag
 lookups during tokenization.
This change also makes the encoder iDCT clear the input buffer like the
 decoder, which can be re-used for the next block, avoiding the need for a
 memcpy or memset in the tokenizer.
This gives a 1.3% speed-up at the default speed-level (1), and a 3.1% speed-up
 at speed-level 2 (for 480p, on x86-64).

svn path=/trunk/theora/; revision=17728