[Python] discrepancy of of schema and schema.metadata if second same-name key is byte-encoded

```python

schema = pa.schema(
[('col1', pa.int8())],
metadata={
'key': 'key provided as string',
b'key': 'key provided as bytes'
}
)
print(schema)
# col1: int8
# -- schema metadata --
# key: 'key provided as string'
# key: 'key provided as bytes'
print(schema.metadata)
# {b'key': b'key provided as string'}
```

**Reporter**: [Chris](https://issues.apache.org/jira/browse/ARROW-17626)
#### Original Issue Attachments:
- [pa_bug.py](https://issues.apache.org/jira/secure/attachment/13048995/pa_bug.py)

_{**Note**: *This issue was originally created as [ARROW-17626](https://issues.apache.org/jira/browse/ARROW-17626). Please see the [migration documentation](https://github.com/apache/arrow/issues/14542) for further details.*}

Start with the attached pa_bug.py reproducer and the pa.schema entry point. Compare how mixed string and byte keys are represented in the printed schema and in schema.metadata, then identify the expected consistent behavior. Done means the discrepancy is resolved and the reproducer is covered by a regression test.

Written by the indexing model from the issue text.