Underfed 0.3.0

• A source that starves in the middle of a session is caught sooner. The ingest is now the
  difference of reservoarr's in_total over the last 45 seconds instead of its in, which is
  a two-minute average: after a sharp drop in needed about 50 seconds to fall under 70% of
  crate, so Dispatcharr's own failover got there first, 105 seconds after the drop on the
  test bench. With 0.3.0 Underfed switched the same feed, dropped to 30%, 79 seconds after the
  drop: the cushion takes about 25 of them to run out, the confirmation 45. in still decides for the first 30 seconds of a feed, after the
  counter goes back (a new reservoarr process) and on lines without the counter; the
  journal's measure says which of the two decided.
• On the reference evening of 7-8 September the triggers stay on the four starved feeds at
  Stable after 0 (45, was 44) and reach one more at the default 180: 272355 at 17:57, which had
  the cushion empty and fed the player 1.64 of 4.6 Mbps while in still read 3.6 (79, was 58).
  On every telemetry line since 30 August the first trigger of each known episode comes
  between 0 and 894 seconds sooner: 336 seconds on Sky Sport 252 on 10 September, 272 on
  202121 on 8 September.
• A channel created after the watcher last read the chains is looked up again when it
  starves, at most once a minute. Before, it was skipped with "no source after this one"
  for up to 15 minutes.

Commits:

• f51f60a CHANGELOG 0.3.0: the test bench numbers instead of the estimate
• 47f4de5 Look the chains up again when a starving channel is not in them
• bf24923 Read ingest off the in_total difference over 45 s instead of the 120 s average