Image dtype of rank filters

[tonysyu]

The rank filters operate on uint8 and uint16 images only. We should allow the user to pass in any dtype and convert if necessary, as suggested in the user guide.

Also, only the first 12 bits of uint16 inputs are used. Wouldn't it make sense to use the largest 12 bits instead? (That way, we could just use the output of img_as_uint.) That said, I'm not sure how to go about changing the behavior. The most obvious change would be to check for values greater than 4095 (as is done now) and do some bit-shifting if values are larger (i.e. a real uint16 value) and issue a deprecation warning if all values are smaller (i.e. a uint12 value put into uint16 array).