Large TGAs wouldn't finish decoding in the fuzzer builds within
30 seconds and that seems ridiculous to me, 30s is a *long* time,
so here's trying to make the TGA decoder more efficient.

The main change is separating the RLE decoding from the pixel
conversions, and only figuring out exactly what conversions to do
once per line, not per pixel. That reduces the overhead to a more
manageable level.