decouple FlushStateToDisk

[yuvicc]

Motivation

Chainstate::FlushStateToDisk() currently does four separable things in one
body: it decides which block files can be pruned, flushes block and undo files
and writes the block index, writes the coins cache, and reports runtime
failures through a BlockValidationState& out-param.

This came up during review of bitcoin#35621, where AcceptBlock() had to grow a dummy
BlockValidationState so that flush failures would not be misread as block
validation failures. As per discussion:

passing a dummy state means the API does not express the real intent...A
cleaner approach would be to expose a validation-state-free method

and suggested either a wrapper method or splitting FlushStateToDisk() into
smaller composable helpers.

Worth stating explicitly, FlushStateToDisk() does not report block invalidity but runtime error.

gettxoutsetinfo, scantxoutset, dumptxoutset rpc and the rolled-back snapshot
helper all call ForceFlushStateToDisk() for reading CoinsDB() and need the dirty coins written out first.
Getting that via FlushStateToDisk() also gives them an opportunistic prune, an m_next_write timer reset, a ChainStateFlushed notification, and a possible chainstate compaction.

SyncCoinsToDisk() is FlushBlockFilesToDisk() (with nothing to prune)
followed by FlushCoinsCache().
The block/undo files and the block index are still written first, and
deliberately so that the coins cache may refer to block index entries that are so
far only dirty in memory, and syncing the coins without them could leave the
on-disk chainstate pointing at a best block the on-disk block index does not
contain, failing LoadChainTip() on restart after a crash.

[optout21]

For reference: bitcoin#35646 changes FlushStateToDisk signature to be able to return FatalError only.
(Note: I'm not implying anything about the merits/relationship of the two approaches, only noting the overlap.)

[optout21]

For reference: bitcoin#35646 changes FlushStateToDisk signature

Oh, my bad, I've overlooked that it's the same author! >D Please ignore

[yuvicc]

For reference: bitcoin#35646 changes FlushStateToDisk signature

Oh, my bad, I've overlooked that it's the same author! >D Please ignore

Correct, was experimenting out! I feel bit by bit changes looks more mergeable rather than a full refactor!