Metadata table scans as streams

[rshkv]

This changes the metadata table APIs to have scan() return streams instead of a single RecordBatch.

Context for this is #863 (comment), where we're adding a metadata table for which we need to read multiple files. @liurenjie1024 suggested we better consume those as stream.

I'm splitting out to discuss the API change separately, even though the existing metadata tables snapshots and manifests only read a single file we're just returning async streams of a single record batch.

cc @liurenjie1024 @Xuanwo

[liurenjie1024]

Thanks @rshkv for this pr, exactly what I mean! We should rebase this pr after #872 ?

[rshkv]

@liurenjie1024, rebased.

We don't strictly need this before merging #863. But would be nice to agree on some stylistic "patterns" for async streams (e.g. using try_stream!, having associated per-batch functions or not).

[Xuanwo]

Thank you @rshkv for working on this, let's move.