pyarpack: improve README documentation.

EXAMPLES/PYARPACK/README

@@ -38,6 +38,17 @@ You can also:
 Note:
 
 1. arpack data type (float, double, ...) must be consistent with A/B numpy dtypes (float32, float64, ...).
+   at python side, the data MUST be casted in the EXACT expected type (int32, int64, float, double, ...).
+   otherwise, C++ may not get the data the way it expects them: C++ will not know how to read python data.
+   if you are not sure how data have been passed from python to C++, set arpackSlv.debug = 1 and check out debug traces.
+   in other words, pyarpack users MUST :
+   1.1. create numpy arrays specifying explicitly the type:
+        >> Aij = np.array([], dtype='complex128')
+   1.2. filling numpy arrays casting value on append:
+        >> Aij = np.append(Aij, np.complex128(np.complex( 200.,  200.))) # Casting value on append is MANDATORY or C++ won't get the expected type.
+   1.3. calling the solver flavor which is consistent with the numpy array data type:
+        >> arpackSlv = pyarpackSlv.complexDouble() # Caution: complexDouble <=> np.array(..., dtype='complex128')
+   note: NO data type check can be done at C++ side, the pyarpack user MUST insure data consistency.
 2. sparse matrices must be provided in coo format (n, i, j, Mij), that is, as a tuple where:
    2.1. n is an integer.
    2.2. i, j, Mij are 1 x nnz numpy arrays.