This removes most transaction request logic from net_processing, and
replaces it with calls to a global TxRequestTracker object.

The major changes are:

* Announcements from outbound (and whitelisted) peers are now always
  preferred over those from inbound peers. This used to be the case for
  the first request (by delaying the first request from inbound peers),
  and a bias afters. The 2s delay for requests from inbound peers still
  exists, but after that, if viable outbound peers remain for any given
  transaction, they will always be tried first.
* No more hard cap of 100 in flight transactions per peer, as there is
  less need for it (memory usage is linear in the number of
  announcements, but independent from the number in flight, and CPU
  usage isn't affected by it). Furthermore, if only one peer announces a
  transaction, and it has over 100 in flight and requestable already, we
  still want to request it from them. The cap is replaced with an
  additional 2s delay (possibly combined with the existing 2s delays
  for inbound connections, and for txid peers when wtxid peers are
  available).

Includes functional tests written by Marco Falke and Antoine Riard.

ninja all check-all

./test/functional/test_runner.py p2p_tx_download