Make $&time more precise and improve its time library handling

[jpco]

This PR is a smaller part of #210, which came from the discussion in #154. I'm not sure about the git history; this PR might be best merged squashed.

It's a fairly heavy refactor of $&time which is, I think, quite a bit clearer, and does a better job than before of handling the messy set of time library functions available. It's not perfect -- gettimeofday(3) is listed as "obsolescent" in the man page -- but swapping out functions is easy with the proposed setup.

In part because of the better time library handling, it also reports times in milliseconds (real-time measurements might be in seconds on machines with don't have gettimeofday(3), though I don't think there are a lot of those.)

; es.old -c 'time sleep 0.1'
     0r     0.0u     0.0s	sleep 0.1
; es.new -c 'time sleep 0.1'
 0.101r   0.001u   0.000s	sleep 0.1

I'd still like those other changes from #210 at some point. This is just the part I feel actively impatient about: in my view, es time is almost useless without millisecond precision.

[jpco]

In order to try to have our cake and eat it too, I have wrapped the millisecond-precision part of this behind a build define, -DMILLISECOND_TIME=1. If you build with that flag, then you'll get millisecond-precision times printed from $&time, and if you don't, then es will default to its current behavior.

This doesn't leave the default behavior perfectly backwards-compatible, since it changes some error-handling behavior and might change certain edge cases around printing times, but I believe it's satisfactorily backwards-compatible to merge now.