The link explains the tests ent does for better than I could.

a) od outputs 16-bit octal values (each byte is a 3-digit string in the range 000..377) while the test matches 2-digit strings so digits 0-3 are naturally overrepresented
b) the test is also counting occurences in the offset column, so some low digits (which ones depend on the sample size) are also overrepresented.

Try something like this instead:

How about the imprecisions in the raw values of the accelerometers? Is't the noise there also from quantum origin and therefore trully random?

Would it use less or more battery for the same amount of random bytes you can get from the camera?