Allow buffered input streams

[V0ldek]

Is your feature request related to a problem? Please describe.
Current implementation reads the entire input to a string. This is not production-viable – very large files that we are targeting with all the performance improvements might not fit in memory. A first step would be to enable buffered reading – load a single page worth of input at a time. There are challenges here – it is possible for a single logical query step to span arbitrarily many blocks, e.g. JSON labels can be arbitrarily long.

Describe the solution you'd like
First of all, current implementations heavily rely on raw AlignedSlice data. This should be abstracted behind a buffered input that can yield slices on-demand.

Two, the query engines need to be made aware of this. They currently rely on having all the data available to index into the slice and compare labels. The engines also need to communicate to the classifiers at which point it is safe to stop keeping old input blocks in memory – we always need the entire label before the currently looked-at colon to be buffered, but after we examine it, it can be discarded.

[V0ldek]

This is not closed yet. We need a smarter buffered input for when Mmap is unavailable, one that does not hold the entire input in memory.

[github-actions]

Tagging @V0ldek for notifications