But now you are timing the startup of `sh` as well.

...which (surprisingly) is a measurable amount. About 1 ms for `dash` and a whooping 9.5 ms for `bash` on a relatively slow CPU:

  pi@pi:~ $ time (for x in {1..1000}; do time echo | sleep 0.01; done 2>/dev/null)
  real    0m24.173s

  pi@pi:~ $ time (for x in {1..1000}; do time sh -c 'echo | sleep 0.01'; done 2>/dev/null)
  real    0m25.170s

  pi@pi:~ $ time (for x in {1..1000}; do time bash -c 'echo | sleep 0.01'; done 2>/dev/null)
  real    0m33.643s

1ms is on the order of process creation time, so not very relevant: The measured program itself requires process creation. Most programs run considerably longer than that (otherwise timing would not make much sense).

It's really not more than that: an optimization.