Support reading Parquet float16 data as float32

[adamreeve]

Description

Support for working with float16 data in Polars is still being debated: #7288

In the meantime, it would be useful to be able to at least read float16 data from Parquet files. Eg:

import numpy as np
import polars as pl
import pyarrow as pa
import pyarrow.parquet as pq


table = pa.Table.from_pydict({
    'x': pa.array(np.array([0.0, 0.5, 1.0, 1.5], dtype=np.float16), type=pa.float16()),
})
pq.write_table(table, 'data.parquet')

df = pl.read_parquet('data.parquet')

Using Polars 1.9.0, this fails with:

polars.exceptions.ComputeError: parquet: File out of specification: Invalid thrift: bad data

Ideally this would work and convert the data to float32, which is what happens if you use pl.read_parquet('data.parquet', use_pyarrow=True), but this isn't possible when using pl.scan_parquet with lazy or streaming operations.

[adamreeve]

The ability to write float32 as float16 Parquet might also be useful, but that would probably depend on #17418