RF - test fails from numpy variability

The cast assumptions seem to differ across platform (different for SPARC
and others) and across numpy versions.  The nibabel code promotes
scalars to arrays for scaling, so this casting should be stable and we
don't need to test it.  So - delete failing test, problem solved :)

Author: matthew-brett

doc/source/devel/scaling.rst

@@ -57,9 +57,6 @@ scalefactor and intercept. If ``(intercept, scalefactor)`` are not the identity
 scalefactor`` by making these be at least 1D arrays, so that floating point
 types will upcast in ``R = (A - intercept) / scalefactor``.
 
-See the file ``nibabel/tests/test_cast_assumptions.py`` for tests of the
-predictability of dtype output given dtype input.
-
 The cast of ``R`` to ``S`` and back to ``R_prime`` can lose resolution if the
 types of ``R`` and ``S`` have different resolution.