Was the ffmpeg decoder actually faster on ARM? I was under the impression that it was "faster" because they tweaked it for desktop chips whereas the original code was designed and coded to be fast on mobile ARM chips, specifically those with NEON SIMD (like every iPhone since the 3GS). I think I saw something claiming that ffmpeg code was faster than Google's code on ARM as long as you disabled the ARM specific assembly but that's a somewhat limited victory.
I'd be interested to know how fast a 3rd party WebM decode could go on an iPhone. From the example of Theora it sounds like it can depend on some very arcane details of how the hardware is set up. I guess we'll get some kind of numbers on that when Android 2.3 starts rolling out wider.