Honor explicit Parquet compression and encoding options

[Mr-J-Forest]

Dataset.to_parquet(..., compression="gzip") raises TypeError: ParquetWriter() got multiple values for keyword argument 'compression' on current main. Explicit use_dictionary and column_encoding options fail for the same reason: the writer supplies defaults as named arguments while leaving the caller's values in **parquet_writer_kwargs.

Pop these three options from the forwarded kwargs, using the existing feature-dependent defaults only when an option is absent. This lets callers choose compression and encoding, including explicit None and False values.

The regression tests write to both paths and binary buffers, read the data back, and inspect Parquet metadata for the requested codecs and encodings. They cover global/per-column compression, disabling compression, dictionary selection, and explicit column encodings, along with unchanged defaults.

Validation on Windows, Python 3.12.13, PyArrow 25.0.1:

• Before the fix: 14 new regression cases failed with the duplicate-keyword error; the two default-option controls passed.
• After the fix: python -m pytest tests/io/test_parquet.py -q — 72 passed.
• Ruff lint and format checks passed for both changed files; git diff --check passed.

Reproducer:

from io import BytesIO
from datasets import Dataset

dataset = Dataset.from_dict({"value": [1, 2, 3]})
dataset.to_parquet(BytesIO(), compression="gzip")