Use zero-copy C data interface in Parquet adapter

[arhamchopra]

We previously used the Arrow IPC interface in the ParquetAdapter to pass data from Python to C++. This would result in the data getting serialized involving additional compute and data copy. We were not able to use the zero-copy interface before due to lack of a portable zero-copy method for arrow data.

The PyCapsule interface, enables easy zero-copy movement of data through the stable C data interface. This PR resolves #212.

This PR replaces the IPC transfer logic with the PyCapsule interface to pass data from Python to C++ zero-copy.
This also results in some performance improvements:

import csp
from csp.adapters.parquet import ParquetReader
import pyarrow as pa

class MyStruct(csp.Struct):
    ts: datetime

@csp.graph
def ReadArrowTable(tables: object):
    reader = ParquetReader(tables, time_column="ts", binary_arrow=True, read_from_memory_tables=True, NEW=True)
    records = reader.subscribe_all(MyStruct, MyStruct.default_field_map())
    csp.add_graph_output("o", records)


def generate_pyarrow_table(num_rows):
    import pyarrow as pa
    import numpy as np
    from datetime import datetime, timedelta

    random_data = np.random.rand(num_rows)
    start_time = datetime(2023, 1, 1)
    timestamps = [start_time + timedelta(seconds=i) for i in range(num_rows)]

    timestamp_array = pa.array(timestamps)
    fields = [ pa.field('ts', pa.timestamp('s')) ]
    schema = pa.schema(fields)
    table = pa.Table.from_arrays([timestamp_array], schema=schema)
    return table

# Old logic:
# %timeit csp.run(ReadArrowTable, generate_pyarrow_table(1000000), starttime=datetime(2023,1,1))
# 1.43 s ± 40.4 ms per loop

# New logic:
# %timeit csp.run(ReadArrowTable, generate_pyarrow_table(1000000), starttime=datetime(2023,1,1))
# 1.25 s ± 20.3 ms per loop

[AdamGlustein]

Looks good to me, wait for @robambalu to take a look as well