fix: Read partitioned parquet files from relative paths

[jeffkinnison]

Fix for #3440. Parquet reads fail under the following conditions:

1. The parquet data is partitioned into multiple files
2. A relative path is passed to pyarrow.parquet.read_table
3. A filesystem object is passed to pyarrow.parquet.read_table

This error can be addressed in two ways:

a. Convert the relative path to an absolute path
b. Do not pass a filesystem object

While determining the root cause, I this StackOverflow question and this pyarrow Jira issue, both of which indicate that there are issues with passing the filesystem object. Both indicate that there are potential issues with passing a filesystem object to pyarrow, so I decided to go for fix b.

This fix modifies ludwig.data.dataset.ray.read_remote_parquet to catch pyarrow.lib.ArrowInvalid errors and try the read again with no filesystem kwarg. It also adds a test for this issue.

[ksbrar]

awesome 🚀

[github-actions]

Unit Test Results

  6 files  ±0    6 suites  ±0   1h 4m 11s ⏱️ - 2m 55s
34 tests ±0  29 ✔️ ±0    5 💤 ±0  0 ❌ ±0 
88 runs  ±0  72 ✔️ ±0  16 💤 ±0  0 ❌ ±0 

Results for commit 40d454d. ± Comparison against base commit 079429e.

♻️ This comment has been updated with latest results.