Dont recompute post on revert <<redux>>

[dirkmc]

Supersedes #3833
Fixes #3560 #3561

The previous state machine implementation has an issue whereby:

• Let's say we're in deadline 5
• We generate a proof for deadline 5 and submit it successfully
• When we reach the challenge epoch for deadline 6 (10 minutes before deadline 6 opens), we start proving for deadline 6
• The chain reorgs such that the revert goes back into deadline 4
  At this point we should resubmit the proof for deadline 5. But the state machine is a single thread, so it would have to either
  • Cancel proving for deadline 6
  • Wait for proving for deadline 6 to complete, then try to resubmit for deadline 5 (if it hasn't already expired)

This PR instead creates two separate go routines:

1. Go routine that just generates proofs:
   Each time the head advances

• If a deadline has expired, abort any ongoing proof for that deadline
• If we reach the challenge epoch for a deadline, start proving
• Once the proof completes, add it to a cache

2. Go routine that submits proofs:
   Each time the head changes

• If the head reverted, invalidate any submitted proofs affected by the revert
• If the head reaches submit confidence for a deadline, and the proof has been generated, submit the proof

[magik6k]

Looks really good, I don't see anything obviously wrong

The only thing to do here is to convert some things to use the now actor abstraction layer

[magik6k]

Looks good, but I'd like to run this on some real miner node before merging

[magik6k]

(Just worked on 3 partitions on my miner)