wrap SqliteCodebase.syncInternal in an immediate transaction

[aryairani]

I was thinking that sqlite could keep more work in RAM if it knows that it's expected to be an atomic operation, and it seems it can. Syncing base was 39% faster with this.

Loose ends

I wondered if PRAGMA read_only would help too, and I stop trying when I got an error about not being able to "prepare" the statement PRAGMA read_only = ? for some reason. You could just have constant statements for = TRUE" and = FALSE"` but I stopped trying. One could also just ask the sqlite mailing list if this would speed things up, they seem to be very knowledgeable.

Timing data

Using this transcript, where base_v2 is a local clone of unisonweb/base@de5f04d and not in the ucm git cache, and repo is an empty repo.

```ucm:hide
.> builtins.merge
.> builtins.mergeio
.> pull /root/base_v2:.trunk base
.> fork base newbase
.> merge builtin newbase
.> cd newbase
.newbase> push /root/unison-slim/repo
.newbase> delete.term Either.mapRight#jruvg9gh6q
.newbase> delete.term List.map#j1ejquc7so
.newbase> push /root/unison-slim/repo
```

Pre-PR (avg of 3 runs)

stage	time
pull	14.05 s
push 1	14.37 s
push 2	0.85 s

With transaction on dest DB (avg of 3)

stage	time
pull	8.99 s (-36%)
push 1	9.27 s (-36%)
push 2	0.85 s

With transaction on src and dest DBs (avg of 7)

stage	time
pull	8.53 s (-39%)
push 1	8.90 s (-38%)
push 2	0.68 s (-20%)

[aryairani]

CI is failing; I'll look into it later.

[pchiusano]

Sweet!