AttributeError: 'ParquetFile' object has no attribute '_columns_dtype' when reading files without pandas metadata

[piotrb5e3]

Describe the issue:
I'm getting the following crash when I try to open a parquet file that was created with either pyspark or pyarrow.

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/piotr.bakalarski/repos/neural-search-commons/venv/lib/python3.10/site-packages/fastparquet/api.py", line 407, in iter_row_groups
    df = self[i].to_pandas(filters=filters, **kwargs)
  File "/Users/piotr.bakalarski/repos/neural-search-commons/venv/lib/python3.10/site-packages/fastparquet/api.py", line 765, in to_pandas
    df, views = self.pre_allocate(size, columns, categories, index, dtypes=dtypes)
  File "/Users/piotr.bakalarski/repos/neural-search-commons/venv/lib/python3.10/site-packages/fastparquet/api.py", line 796, in pre_allocate
    dtypes, self.tz, columns_dtype=self._columns_dtype)
AttributeError: 'ParquetFile' object has no attribute '_columns_dtype'

Minimal Complete Verifiable Example:

Using fastparquet==2023.4.0 and pyarrow==10.0.1:

import pyarrow.parquet as pq
import pyarrow as pa
from fastparquet import ParquetFile

pq.write_table(pa.table({"a": "abcd", "n": [1,2,3,4]}), "repro.parquet")
pf = ParquetFile("repro.parquet")

pf.head(1)  # <-- This line causes error

Anything else we need to know?:
From what I can see, this is caused by _columns_dtype not getting set, because the __setstate__ in __getitem__ [HERE] does not set _columns_dtype. Overall it seems that this getitem behavior is unintuitive and may lead to more issues with not-copied parameters.

Environment:

• Dask version: None
• Fastparquet version: 2023.4.0
• Python version: 3.10.11
• Operating System: macOS Ventura 13.4.1
• Install method (conda, pip, source): pip

[piotrb5e3]

I can see that this was already fixed in Extra field when cloning ParquetFile (#866). Could you release a new version of the library?