Fix compatibility quirks between arrow and parquet structs

[nevi-me]

Describe the bug

See #246 and 6a65543. There are some notes referring to this issue in that PR.

The issue is that the different parquet implementations handle non-null structs (and possibly lists) differently.
Spark doesn't seem to have a facility to create non-null struct schemas, so structs are nullable by default. If one creates a non-null struct with null children, pyspark won't read it.

The C++ implementation reads this back fine, perhaps because there's a good mapping to Arrow data.
The Rust implementation will write the file, but won't read it back.

I also have some uncertainty on whether a non-null parent + null child is logically correct or Arrow specification compliant.

To Reproduce

• Create a RecordBatch that has a non-null struct with a nullable child.
• Write that to Parquet
• Read the Parquet file with Spark

Expected behavior

There shoulb some clear behaviour that is also documented.

Additional context

See the commit 6a65543, specifically the comments added around the tests.