make sure header PMMR is init correctly based on header_head from db

[antiochp]

We had logic in chain init to make sure header_head was consistent with the header PMMR (in the case where it was missing from the db after migration to 3.x.x).

This PR adds logic to ensure the inverse of this - that the header PMMR is consistent with header_head from the db in the case where the header PMMR contains unwanted rightmost data.
The PMMR is append-only and there may be headers appended that we want to discard.
The header_head in the db is our "source of truth".

We simply (re)set the PMMR handle last_pos to discard data to the right.
Then we can safely "rewind" the header PMMR to the pos corresponding header_head from db.

---

Also tweaked the order of operations slightly when processing block headers.
We now validate the header before applying it to our header PMMR (to validate the roots).
Otherwise we potentially leave the header PMMR in a bad state if the roots are valid but the header ends up being invalid for some other reason.
This is likely the reason why a node gets into a bad state. We can now recover from this on startup but we want to avoid it happening in the first place.

---

Tested with sync (and staying in sync) on both floonet and mainnet.
Tested mining on usertesting.
@quentinlesceller tested this against floonet with stratum miners.