By data path, I mean the width of the values that are read from the stacks, program memory, and device bus. Pairs of 8-bit values can be treated as 16-bit values in order to perform wider arithmetic, but all data ultimately moves around the system as 8-bit values.
Whether 16-bit data moves around the system as 8-bit values or not sounds like a question about the implementation, not the architecture (the spec).
For example, the spec says, "Reading a double from program memory will read the high byte of the double from the given address and the low byte from the following address," but I'd think that generally you'd want the implementation to work by reading the whole 16-bit word at once and then byte-swapping it if necessary, because that would usually be faster, and there's no way for the program to tell if it's doing that, unless reading from the first byte has a side effect that changes the contents of the second byte or otherwise depends on whether you were reading the second byte at the same time.
(Of course if you have a "double" that crosses the boundaries of your memory subsystem's word, you have to fall back to two successive word reads, but that happens transparently on amd64 CPUs.)
