[BugFix] Robustize cuFINUFFT Python guru interface.

As per the docs:
* finufft.Plan(n_modes_or_dim) accepts (int, tuple[int])
* cufinufft.Plan(n_modes) accepts (int, tuple[int])

In practice `n_modes_or_dim` is transformed internally to a NumPy array, hence providing any iterable to finufft.Plan() works.
However `n_modes` in cufinufft.Plan() is more restrictive: it works with tuple[int], fails planning if list[int], and wrongfully passes planning if ndarray[int], but then seg-faults at execute().

This commit increases the robustness of cufinufft.Plan() to match the flexibility of the CPU interface.

python/cufinufft/cufinufft/_plan.py

@@ -5,6 +5,8 @@
 """
 
 import atexit
+import collections.abc
+import numbers
 import sys
 import warnings
 
@@ -108,8 +110,12 @@ def __init__(self, nufft_type, n_modes, n_trans=1, eps=1e-6, isign=None,
         else:
             raise TypeError("Expected complex64 or complex128.")
 
-        if isinstance(n_modes, int):
+        if isinstance(n_modes, numbers.Integral):
             n_modes = (n_modes,)
+        elif isinstance(n_modes, collections.abc.Iterable):
+            n_modes = tuple(n_modes)
+        else:
+            raise ValueError(f"Invalid n_modes '{n_modes}'")
 
         self.dim = len(n_modes)
         self.type = nufft_type