pyon: Correctly deserialize bare NaNs

This also fixes (non-numpy) lists containing NaNs.

Previously, accidentally storing a NaN in a dataset would
bring down large parts of the system.

artiq/protocols/pyon.py

@@ -194,6 +194,7 @@ _eval_dict = {
     "false": False,
     "true": True,
     "slice": slice,
+    "nan": numpy.nan,
 
     "Fraction": Fraction,
     "OrderedDict": OrderedDict,

artiq/test/test_serialization.py

@@ -26,6 +26,8 @@ class PYON(unittest.TestCase):
             with self.subTest(enc=enc):
                 self.assertEqual(pyon.decode(enc(_pyon_test_object)),
                                  _pyon_test_object)
+                # NaNs don't compare equal, so test separately.
+                assert np.isnan(pyon.decode(enc(np.nan)))
 
     def test_encdec_array(self):
         orig = {k: (np.array(v), np.array([v]))