feat: Implement shared delete file loading and caching for ArrowReader

[dojiong]

Which issue does this PR close?

• Closes #.

What changes are included in this PR?

Currently, ArrowReader instantiates a new CachingDeleteFileLoader (and consequently a new DeleteFilter) for each FileScanTask when calling load_deletes. This
results in the DeleteFilter state being isolated per task. If multiple tasks reference the same delete file (common in positional deletes), that delete file is
re-read and re-parsed for every task, leading to significant performance overhead and redundant I/O.

Changes

• Shared State: Moved the DeleteFilter instance into the CachingDeleteFileLoader struct. Since ArrowReader holds a single CachingDeleteFileLoader instance across
  its lifetime, the DeleteFilter state is now effectively shared across all file scan tasks processed by that reader.
• Positional Delete Caching: Implemented a state machine for loading positional delete files (PosDelState) in DeleteFilter.
  • Added try_start_pos_del_load: Coordinates concurrent access to the same positional delete file.
  • Added finish_pos_del_load: Signals completion of loading.
  • Synchronization: Introduced a WaitFor state. Unlike equality deletes (which are accessed asynchronously), positional deletes are accessed synchronously by
    ArrowReader. Therefore, if a task encounters a file that is currently being loaded by another task, it must asynchronously wait (notify.notified().await)
    during the loading phase to ensure the data is fully populated before ArrowReader proceeds.
• Refactoring: Updated load_file_for_task and related types in CachingDeleteFileLoader to support the new caching logic and carry file paths through the loading
  context.

Are these changes tested?

Added test_caching_delete_file_loader_caches_results to verify that repeated loads of the same delete file return shared memory objects

[mbutrovich]

Thank you @dojiong! I will take a look since I worked on this code recently. I'll also test the changes with Comet and the Iceberg Java tests.

[mbutrovich]

I ran Iceberg Java 1.10.0 tests with Comet against this branch:

iceberg-spark:iceberg-spark-extensions-3.5_2.13:test:

iceberg-spark:iceberg-spark-3.5_2.13:test

I did a quick review of the code, and I will take another read through later today. I mostly just need another cup of coffee before I think too hard about race conditions with the Notify usage.

[mbutrovich]

This LGTM, just a (possibly sound) performance suggestion. Thank you @dojiong!

[liurenjie1024]

Thanks @dojiong for this fix, and Matt for review. Let's wait for a while befor this got merged.