Migrate caching of compressed data from SpanManager to BlobReader

[vkuzniet]

Is your feature request related to a problem? Please describe.
Once #231 is complete, SpanManager needs to stop having the compressed spans cache and keep caching only uncompressed spans going forward. Compressed spans should be cached within BlobReader. Please refer to the design doc from #103 for more details on the concepts.
In short, we need to do the following:

1. BlobReader now has the cache of spans (aka regions).
2. SpanManager keeps only the cache of uncompressed spans. If it needs the compressed span, it requests it from BlobReader.

Describe the solution you'd like
The outcome for this issue is a design for the separation.

Describe alternatives you've considered
Keep everything as is.

Additional context
#103

[Kern--]

This needs additional design work. I will update this comment with more specific information later.

[Kern--]

While we're doing this work, we should be thinking about the ability to decouple the actual network requests from the span size. A span is really a unit of decompression, but we also use it as a unit of data to be fetched from the network. We can probably get some better performance if we optimize our network requests separately from the decompression. E.g. S3 likes 8 or 16MiB requests, but we might want 2 or 4MiB spans to reduce necessary compute.

We don't actually have to solve this problem in the initial separation, but we shouldn't design it in such a way that we can't do that later.

[Kern--]

TODO: Merge with #534