refactor: increased code sharing between CPU and GPU interpretation in RNTuple reading

[fstrug]

Much of the code used to read RNTuple data via cpu and gpu interpretation is common but contained in separate functions. This means any changes made in one workflow will not be automatically reflected in the other. This pr increases function sharing between the two interpretation modes such that any code improvements are automatically shared. A FieldClusterMetadata class has been added to contain metadata used in the reading and deserializing of raw page data.

I have also changed the arrays() argument use_GDS to interpreter since it is not strictly necessary to have GDS support for gpu interpretation of RNTuple data and is more reflective of what the argument is controlling.

[ianna]

@fstrug - nice! Please check some minor comments. Thanks.

[ariostas]

Thank you, Frank! This looks great to me. I like how you factored out some pieces to make the code cleaner.

The only comment I have is that I'm not very sure about the interpreter keyword argument, since I think it's a bit ambiguous. It might make sense to implement decompression_executor and interpretation_executor like how is done for TTrees, but I'm not familiar with those, so I'd have to look into that. So that can just be a follow-up.

At some point we'll have to sit down and fix all the keyword arguments for the common methods (arrays, iterate, etc). But I think it's still okay if we change some of the lesser used keyword arguments while we figure things out and people are not regularly working with RNTuples.