regression: microdescriptor performance

[meejah]

As part of "switching to Automat for state-machines", the microdescriptor parsing state-machine was switched over. This, however, has made TorState start-up times very slow.

[meejah]

To quantify this a little, on Python3:

• pre-Automat: 22000 microdescriptors/second
• master: 2400 microdescriptors/second
• master + some Automat hacks: ~9000 microdescriptors/second

[meejah]

on pypy, this is:

• pre-automat: ~76000
• master: ~6000
• master + tweaks: ~27000

[meejah]

(of interest: stem's microdescriptor parser does ~7300/s on py2 and ~8700/s on py3)

[meejah]

I'm marking this for 0.19.0 -- I want at least revert to previous (or different) code for the microdescriptor parser before the next release. Building TorState objects has gone from "dozens of milliseconds" to "a couple seconds" which is very noticeable and way too slow for this.

[meejah]

For the 0.19.0 release, I put back the old parser code (using the spaghetti.FSM state-machine code). See ba96e8ba167882ecce4fdb3789701a723c9a8f7c