Rewrite lingering.

The two major changes here are:

- Move lingering out of NetworkMonitor. The fact that lingering
  is currently its own state in NetworkMonitor complicates the
  logic there: while a network is lingering it cannot be in any
  other state, we have to take care not to leave LingeringState
  for the wrong reason, etc.
- Instead of keeping a single per-network boolean to indicate
  whether a network is lingered or not, keep a linger timer for
  every request. This allows us to fix various corner-case bugs
  in lingering.

The changes in behaviour compared to the current code can be seen
in the unit test changes. Specifically:

1. Bug fix: when a network is lingered, and a request is added
   and removed to it, the existing code tears the network down
   immediately. The new code just sends another CALLBACK_LOSING
   and resumes lingering with the original timeout.
2. Bug fix: if cell is unvalidated and wifi comes up and
   validates before cell does (as might happen on boot), the
   existing code immediately tears down cell. The new code
   lingers cell, which is correct because unvalidated cell was
   the default network, so an app might have been using it.
3. Correctness improvement: always send CALLBACK_AVAILABLE for
   the new network before sending CALLBACK_LOSING. This was not
   really an issue in practice, because the usual flow is:
    - Network A is the default.
    - Network B connects, CALLBACK_AVAILABLE.
    - Network B validates, CALLBACK_LOSING.

Bug: 23113288
Change-Id: I2f1e779ff6eb869e62921a95aa9d356f380cf30a