Add custom MultiFileReader for reading delete files

[mchataigner]

Context: Slower read performance when a table has many delete files.

TL;DR: We can leverage the metadata already available in DuckLake to improve load time of delete files.

Problem & Motivation:

DuckLake stores file_size metadata for both data and delete files. For data files, there is already a mechanism to forward this metadata to the MultiFileReader and the underlying filesystem. The Parquet reader requires this file_size to access the footer metadata. When using an HTTPFileSystem instance (e.g., for S3, Azure), it performs a HEAD request on the file if metadata fields (file_size, etag, last_modified) are not present. Since all files in DuckLake are immutable, we can apply the same optimization logic for delete files to avoid these unnecessary HEAD requests.

Solution:

Implements a custom multi-file reading solution that pre-populates file metadata to eliminate redundant storage HEAD requests when scanning delete files:

Key Changes:

1. New DeleteFileFunctionInfo struct: Extends TableFunctionInfo to carry DuckLakeFileData metadata through the table function binding process.

2. Custom DeleteFileMultiFileReader class:

   • Extends DuckDB's MultiFileReader to intercept file list creation
   • Pre-populates ExtendedOpenFileInfo with metadata already available from DuckLake:
     • File size (file_size_bytes)
     • ETag (empty string as placeholder)
     • Last modified timestamp (set to epoch)
     • Encryption key (if present)
   • Creates a SimpleMultiFileList with this extended info upfront
   • Overrides CreateFileList() to return the pre-built list, bypassing DuckDB's default file discovery

3. Modified ScanDeleteFile() method:

   • Changed parquet_scan from const reference to mutable copy to allow modification
   • Attaches DeleteFileFunctionInfo and custom reader factory to the table function
   • Passes the actual parquet_scan function to TableFunctionBindInput instead of a dummy function, ensuring proper function context

Performance Impact: Eliminates HEAD requests to object storage when opening Parquet delete files. This is particularly beneficial when working with remote storage (S3, Azure, etc.) and tables with many delete files, where HEAD requests were causing significant performance bottlenecks.

[pdet]

Hi @mchataigner thanks for the PR!
Could you add a MinIO test that demonstrates fewer requests are done ?
Could you also retarget it to v 1.4?

[mchataigner]

@pdet you're welcome. I will do the changes.
Thanks for your feedback.

[mchataigner]

@pdet sorry for the delay, I updated the PR with format fix and added a test with MinIO.