publish: build checkpoint files during ledger close

[marta-lokhova]

Resolves #4212
Resolves #4469

The change removes publish dependency on txhistory and ledgerheaders SQL tables, and instead constructs ledger headers/tx sets/tx results checkpoints at ledger close. This allows us to deprecate and drop support for txhistory entirely, which is beneficial for performance. Additionally, the code became simpler, and the xdr->sql->xdr roundtrip processing has been removed (as part of this, txsethistory table was dropped as well). The tradeoff in this approach is that we lose atomicity we had with SQL transactions. As a result, core now manually has to handle crashes. This is done by always constructing "tmp" checkpoints first, and only promoting them after a checkpoint ledger has been committed to the database. In the event that tmp file has been written, but core crashed in apply and rolled back the transaction, there's automatic file cleanup on startup, that trims any uncommitted ledgers from the temp file.

This change also drops publishqueue SQL table completely in favor of checkpoint files containing HAS. A similar recovery mechanism is used to cleanup files in case of a crash, except it's simpler, since we can just delete whole checkpoint files instead of truncating them.

Note: now that we drop txhistory, we can also deprecate the --force-back flag.

[marta-lokhova]

Note: the current version of this PR drops txhistory table for simplicity, however depending on who depends on it, we might need to still support it, while also building checkpoint files directly.

[graydon]

Broadly speaking: great work! Handful of nits, handful of clarifying questions / requests for documentation about what's going on, and a couple places I'm uncertain about the recovery logic that I'd like to tighten up a bit (or understand what I'm currently failing to). But overall happy to see, and happy to collaborate on wrapping it up tomorrow.