Pre-hash rows and add SmallHashSets

[kazimuth]

Description of Changes

Annotate rows with their hashes. This is halfway to the old solution of annotating code with their serialized forms; it speeds up insertion into BTreeIndexes while saving a byte array allocation per-row.
(Wait, why do we hash when inserting into BTreeIndexes, you ask? It's because we use a HashSet to store the multiple rows corresponding to a non-unique key.)

This saves work on the main thread repeatedly hashing values. Once we parallelize message processing, this work will be spread out over multiple threads too.

Also adds a data type SmallHashSet<T> for use in BTreeIndexes. This is a struct that can store at most one element without allocating. This gives dramatic performance improvements on initial connection -- it seems that most rows in BTreeIndexes don't have any other rows with the same key. So, skipping all allocations in this case gives very good performance.

API

• This is an API breaking change to the SDK

Requires SpacetimeDB PRs

Testsuite

SpacetimeDB branch name: master

Testing

I will test:

• blackholio
• bitcraft

[kazimuth]

This does mildly pessimize calls to Filter -- it adds a second allocation to them.
SmallHashSet is so good I'm tempted to just use it and get rid of PreHashedRow, not sure.