Support python-style negative indices in axes arguments

In python, indexing as array[-k] is the same as array[N-k]. This extends that behavior to the axes argument.

pypocketfft.cc

@@ -63,13 +63,18 @@ shape_t makeaxes(const py::array &in, py::object axes)
       res[i]=i;
     return res;
     }
-  auto tmp=axes.cast<shape_t>();
-  if ((tmp.size()>size_t(in.ndim())) || (tmp.size()==0))
+  auto tmp=axes.cast<std::vector<ptrdiff_t>>();
+  auto ndim = in.ndim();
+  if ((tmp.size()>size_t(ndim)) || (tmp.size()==0))
     throw runtime_error("bad axes argument");
-  for (auto sz: tmp)
-    if (sz>=size_t(in.ndim()))
-      throw runtime_error("invalid axis number");
-  return tmp;
+  for (auto& sz: tmp)
+    {
+    if (sz<0)
+      sz += ndim;
+    if ((sz>=ndim) || (sz<0))
+      throw invalid_argument("axes exceeds dimensionality of output");
+    }
+  return shape_t(tmp.begin(), tmp.end());
   }
 
 #define DISPATCH(arr, T1, T2, T3, func, args) \