Use pyarrow.field to preserve nullability in Arrow conversion (#602)

* Add nullability flags throughout ak.to_arrow.

* Using nullability flags in ak.from_arrow.

* Streamlined handling of lists and added ak.to_arrow_table.

* Non-nullability passed through Parquet, and Arrow 2.0 can convert records in nested lists through Parquet.

* Fix for Python 2.

* More Python 2 fixes.

src/awkward/__init__.py

@@ -33,21 +33,9 @@
 # third-party connectors
 import awkward._connect._numpy
 import awkward._connect._numba
-
-numba = type(awkward._connect._numba)("numba")
-numba.register = awkward._connect._numba.register
-
 import awkward._connect._numexpr
-
-numexpr = type(awkward._connect._numexpr)("numexpr")
-numexpr.evaluate = awkward._connect._numexpr.evaluate
-numexpr.re_evaluate = awkward._connect._numexpr.re_evaluate
-
 import awkward._connect._autograd
 
-autograd = type(awkward._connect._autograd)("autograd")
-autograd.elementwise_grad = awkward._connect._autograd.elementwise_grad
-
 # high-level interface
 behavior = {}
 from awkward.highlevel import Array

src/awkward/_connect/_autograd.py

@@ -2,6 +2,8 @@
 
 from __future__ import absolute_import
 
+import types
+
 import awkward as ak
 
 numpy = ak.nplike.Numpy.instance()
@@ -83,4 +85,5 @@ def getfunction(inputs):
     return broadcast
 
 
-elementwise_grad.elementwise_grad = elementwise_grad
+ak.autograd = types.ModuleType("autograd")
+ak.autograd.elementwise_grad = elementwise_grad

src/awkward/_connect/_numba/__init__.py

@@ -3,6 +3,7 @@
 from __future__ import absolute_import
 
 import distutils.version
+import types
 
 import awkward as ak
 
@@ -108,3 +109,7 @@ def castint(context, builder, fromtype, totype, val):
         return builder.trunc(val, context.get_value_type(totype))
     else:
         return val
+
+
+ak.numba = types.ModuleType("numba")
+ak.numba.register = register

src/awkward/_connect/_numexpr.py

@@ -5,6 +5,7 @@
 import warnings
 import sys
 import distutils.version
+import types
 
 import awkward as ak
 
@@ -153,4 +154,6 @@ def getfunction(inputs):
     return ak._util.wrap(out[0], behavior)
 
 
-evaluate.re_evaluate = re_evaluate
+ak.numexpr = types.ModuleType("numexpr")
+ak.numexpr.evaluate = evaluate
+ak.numexpr.re_evaluate = re_evaluate