Refactor read_csv to support converters (#993)

* Refactor read_csv to support converters

Motivation: support pd.read_csv use with converters to avoid costly
unboxing of string columns read with pyarrow read_csv when returning
DF from objmode.

* Adds StdStringView Numba type to hstr_ext

Motivation: for optimization purposes (avoiding copy when creating NRT manageble
unicode instances) when working with string data stored in native extensions.

* Adds zip and dict builtins overloads to support easy literal dict ctor

Motivation: there's no easy way to create Numba LiteralStrKeyDict
objects for const dicts with many elements. This adds a special overload
for dict builtin that creates LiteralStrKeyDict from tuple of pairs
('col_name', col_data).

* Replacing zip overload builtin with internal sdc_tuple_zip function

Details: zip builtin is already overloaded in Numba and has priority
over user defined overloads, hence in cases when we want zip two single
elements tuples, e.g. zip(('A', ), (1, )) builtin function will match
and type inference will unliteral all tuples, producing iter objects
(that are always homogeneous in Numba). That is, literality of objects
will be lost. Using sdc_zip_tuples explicitly avoid this problem.

* Fixing issue with literal dict ctor with single element

* Moving stringlib to native

* Fixing refcnt issue and adding tests

* Adding rewrite for dict(zip()) calls

* Fixing str_view_to_float impl and tests

* Fixing refcnt problem with pyarrow table ptr

* Fixing bugs found in failed tests and examples

Author: kozlov-alexey

sdc/datatypes/hpat_pandas_functions.py

@@ -72,6 +72,7 @@
 from sdc import sdc_autogenerated
 from sdc.functions import numpy_like
 from sdc.hiframes.api import isna
+from sdc.hiframes.join import setitem_arr_nan
 from sdc.datatypes.hpat_pandas_groupby_functions import init_series_groupby
 from sdc.utilities.prange_utils import parallel_chunks
 from sdc.extensions.indexes.indexes_generic import sdc_indexes_join_outer, sdc_fix_indexes_join
@@ -679,7 +680,7 @@ def sdc_pandas_series_setitem_idx_bool_array_align_impl(self, idx, value):
                             if self_index_value in map_index_to_position:
                                 series_data[i] = value._data[map_index_to_position[self_index_value]]
                             else:
-                                sdc.hiframes.join.setitem_arr_nan(series_data, i)
+                                setitem_arr_nan(series_data, i)
 
             else:
                 # if value has no index - nothing to reindex and assignment is made along positions set by idx mask
@@ -734,7 +735,7 @@ def sdc_pandas_series_setitem_idx_bool_series_align_impl(self, idx, value):
                         value_index_pos = map_value_index_to_position[idx_index_value]
                         self._data[self_index_pos] = value._data[value_index_pos]
                     else:
-                        sdc.hiframes.join.setitem_arr_nan(self._data, map_self_index_to_position[idx_index_value])
+                        setitem_arr_nan(self._data, map_self_index_to_position[idx_index_value])
             else:
                 # use filtered index values to create a set mask, then make assignment to self
                 # using this mask (i.e. the order of filtered indices in self.index does not matter)