Further optimize BlobDataProvider constructor

[sffc]

In #2603, I improved the performance of the BlobDataProvider constructor by 60%. However, there is still a lot more that can be done. In principle, this should be a constant-time operation, except that we need to do minimal validation of the bytes.

There are two parts of the constructor that take the most time:

1. Checking that the indices list in the VarZeroVec is sorted (a bigger chunk)
2. Checking that the joiner list in the ZeroMap2d is sorted (relatively smaller)

Both of these get slower as the data file gets bigger, especially (1).

If we could skip this during construction and delay the check to data load time, we should be able to get the BlobDataProvider construction time down to 100 ns or less.

I did one attempt at this in #2609 by changing the data model. However, this can be done without touching the data model, meaning we can do this later and still read the same data files.

CC @zbraniecki

[robertbastian]

Uhm I'm not sure why this is assigned to me. Is there an AI?

We could have a GIGO version of ZeroMap that doesn't check the sorting on deserialization.

[sffc]

I think when I assigned this issue I thought it was something different. I'll take it again.