Optimized the NaN box to be byte aligned, performance under the fibtest.lua script improved by ~2.5 seconds (~31 before ~28.5 after on cpunch's machine) also cleaned up some misc. comments
