I should have been suspicious of the Gorm/Gorp numbers matching exactly :)

As I had expected, this removes an unfair disadvantage for Oblast, which
had to maintain two differently-typed slices for Insert() and Delete().
This is not really a fair comparison since usually you would not be
doing both operations in one go.

For Gorp, this rewrite is neutral. But in a shocking twist, Gorm somehow
does _significantly_ worse with a specialized N=1 implementation:

```
BenchmarkInsertAndDelete/N=1/via_Gorm-24
before: 11508 B/op    157 allocs/op
after:  12840 B/op    172 allocs/op
```