Regarding performance, we have tested this with Csound and in that system, doubles are actually faster in 64bit systems (but slower, not by much, in 32-bit). Audio quality has reportedly been much higher.
Not necessarily true in FluidSynth, but in theory, doubles would be faster in 64-bit systems (but slower in 32-bit). The more pressing concern is the memory usage, which would be, well, doubled, on both 32-bit and 64-bit systems. This could break on some less memory-intensive systems.
I guess FluidSynth itself should be tested in all of these combinations, but my gut feeling is that switching build systems has enough chance of breaking on various machines that it's best if it doesn't also change the default build parameters. (Or you may end up with a scenario like "Fluid doesn't work / is laggy with Cmake but it worked fine with Autotools," and then spend ages debugging the Cmake build when the problem was a changed build parameter.) So it should probably default to single-precision.
If it's shown through experimentation that doubles are better overall than floats, then the default parameter should be changed in both the autotools and cmake builds at the same time, and noted in the changelog.
Matt Giuca