BigDeedle improvements

[tpetricek]

This makes the addressing mechanism in BigDeedle a bit more flexible (as needed for internal BMC work). In particular:

• We no longer access series length when printing (there is a new operation that lets us take X elements from the front and from the end - printing looks at first 30 elements (to see if there is enough) and then prints first and last 15).
• We no longer assume that the addresses are continuous (this lets us use partitioning where e.g. first 32 bits encode partition index and next 32 bits encode offset within partition). Address is still int64, which we might need to generalize in the future.

NOTE: There is more cleanup to be done here, so no need to merge/review this yet.

[tpetricek]

I guess we should merge this into master at some point!

I have used this branch for the latest Deedle release (for both the use of BigDeedle at BMC and for public NuGet package) and my main reason for not merging this is that there are some places that I would like to clean up and that BigDeedle needs a public demo and documentation.

@adamklein @hmansell What do you think would be the best strategy? Should I write a summary of the changes I did that are affecting "small Deedle" so that you can look at those (and we merge it with BigDeedle work in progress)? Or do you think it is OK to wait a bit more and merge it once it is all nice? (I think I need ~2 days to just make everything a bit nicer).

[hmansell]

Probably making it nicer is better, IMO.

[adamklein]

Looked over the changes, whenever you want to merge this to master, no problems here.

[tpetricek]

OK, I think this is now ready to merge.

I did some more debugging while working on the public BigDeedle demo, which also revealed some interesting issues. The biggest one is that with BlockStore (and the public demo, which follows this pattern), the addresses we use are no longer compatible between multiple different representations of a vector (in-memory uses just index, but the address for BigDeedle means completely different thing).

This was correctly handled in most places, but there were still cases where wrong addressing scheme was used. This is now solved by marking vectors and indices with "addressing scheme" that's checked and enforced to match.

There is some comment about this in the design notes.

I also did a general cleanup of the BigDeedle code, so all the tricky things are now commented and it should actually be understandable & readable :)

[tpetricek]

I'm going to merge this - this just improves the existing PR that @adamklein reviewed and I also improved the documentation and cleaned up the code as I suggested and as @hmansell asked for.

The BigDeedle code is still quite complex, but I think merging it into master and sending other fixes and improvements as smaller PRs will be a better way to continue developing this.

(Also, I'll push a new release with the most recent improvements for the internal BMC needs.)