`OverflowError` when a chunk exceeds 2 GB

[chase-dwelle]

DatasetCoder.estimate_size returns value.nbytes directly (core.py#L275). For
chunks larger than 2**31 - 1 bytes (~2 GB), this overflows the C int in Beam's Cython CallbackCoderImpl.estimate_size and raises:

OverflowError: value too large to convert to int

Noticed this hitting after upgrading to the latest version when loading datasets from GRIBs with multple atmostpheric levels, e.g., ECMWF IFS ensembles (51 members × 13 levels × 721 × 1440), where a single per-key xarray.Dataset chunk exceeds 2 GB before any rechunk.

Minimal repro:

import apache_beam as beam
import numpy as np
import xarray as xr
import xarray_beam as xbeam

ds = xr.Dataset({"x": (("a", "b"), np.zeros((20000, 14000)))})  # ~2.1 GB

with beam.Pipeline() as p:
    _ = (
        p
        | beam.Create([(xbeam.Key({}), ds)])
        | beam.Reshuffle()
    )

Capping the returned estimate at 2**31 - 1 (with a warning so oversized chunks are still visible) avoids the crash, will put up a PR.