Naive interleaved filtering and data reading

[AdamGS]

Read filter-related data using a different LayoutReader, but with a shared message cache. Once we wire-in metadata, we can use this information to skip chunks/parts of chunks.

There are some assumptions on overall alignments/batch size, that are held by the underlying BatchReader.

[robert3005]

This removes having to unify schemas of filters and projections but having two readers is awkward. The scan to layout is not used beyond being used for schema projection and we keep leaning on stream having a copy of top level scan