This adds at least one memory allocation (`collect` into a `Vec`) but is still faster for a reasonable number of IDs.

There is a performance penalty (+13%) but this collection enables cheap
deletions.

When creating an index (`IndexedReader::create_index`) the unused `Blob`s are not copied into memory, which really improves performance for this operation (5x faster).

This allows filtering elements and finding their dependencies by
decoding as few blobs as possible.