Revert "Replace dtype (uint8 or uint16) with dtype int"

I don't think I agree with the commit this reverts
(922710f).

Yes, you can technically pass dtype=np.int64 or similar as an `image`.
But this will incur precision loss and a warning, urging you to convert
to uint16 before.

This seems unexpected in context of dtype int. The reason we are doing
this is so that type checkers can be useful and can catch issue like
this beforehand. They won't if we use int here.

Instead, I think we should treat this as an undocumented backwards
compatible fallback.

Author: lagru

src/skimage/filters/rank/_percentile.py

@@ -122,7 +122,7 @@ def gradient_percentile(
 
     Parameters
     ----------
-    image : ndarray of shape (M, N) and dtype int
+    image : ndarray of shape (M, N) and dtype (uint8 or uint16)
         Input image.
     footprint : ndarray
         The neighborhood expressed as a 2-D array of 1's and 0's.
@@ -167,7 +167,7 @@ def mean_percentile(
 
     Parameters
     ----------
-    image : ndarray of shape (M, N) and dtype int
+    image : ndarray of shape (M, N) and dtype (uint8 or uint16)
         Input image.
     footprint : ndarray
         The neighborhood expressed as a 2-D array of 1's and 0's.