Support Gilbert-Elliott packet-loss model

[nick-e]

Closes #1546.

Replace LinkConditioner's per-packet drop rate with a two-state Gilbert-Elliott Markov model so simulated packet loss occurs in correlated bursts, matching real packet-loss behavior.

Give LinkConditioner a good/bad state. A good state denotes the link is healthy and not dropping packets. A bad state denotes the link is dropping a lot of packets. It also opens up the possibility of adding an extra latency to packets during a bad state.

Add builder methods to LinkConditionerConfig to configure the packet loss model.

[nick-e]

To address the convo in the previous PR:

I didn't change the values for good_condition(), average_condition(), and poor_condition() because I don't know what settings to pick yet. I'm working on capturing real network traffic of connections within a city and across USA's coasts in order to create realistic configs. I need #1549 for that.

[nick-e]

To address the convo in the previous PR:

I didn't change the values for good_condition(), average_condition(), and poor_condition() because I don't know what settings to pick yet. I'm working on capturing real network traffic of connections within a city and across USA's coasts in order to create realistic configs. I need #1549 for that.

I've added some LinkConditionerConfig presets in #1562 (that PR contains the fit() function used to create the presets).

[cBournhonesque]

Thanks, looks great