CP-307958: small DB improvements and benchmarks

[edwintorok]

We were converting events to RPC.t twice: first using the internal OCaml names, and then mapping them back to some "wire" names. This can be done a lot more simply by generating the wire names in the first place (the ppx support overriding the wire name of fields).

ministat confirms an improvement on rpc_of_event:

      N           Min           Max        Median           Avg        Stddev
x  922     440.79126     46094.975     467.83727     563.79259     1676.0843
+ 1131     51.346821     33407.866     58.715393     115.98006     1105.4178
  Difference at 95.0% confidence
            -447.813 +/- 120.966
             -79.4286% +/- 13.1488%
           (Student's t, pooled s = 1390.95)

Atomic is also faster than an integer wrapped with a mutex:

Mutex+incr (ns):
 { monotonic-clock per run = 32.272091 (confidence: 32.365531 to 32.170138);
 r² = Some 0.999441 }
Atomic.incr (ns):
 { monotonic-clock per run = 2.938678 (confidence: 2.944486 to 2.933688);
   r² = Some 0.999857 }

I've added some new new benchmarks for event field conversion, atomic vs mutex and DB field writes.

[last-genius]

There is a bit of weirdness again, probably with jj doing something so that the last commit links to changes from 5 commits instead:

f2dcc67

Looks good otherwise

[lindig]

Looks good. I see this mostly as a move from references to Atomic.