Previous commit 01b95ac from
bitcoin#33212 fixed one case where index state
referring to blocks the node had not flushed to disk yet could be commited,
leading to index corruption if the node was not shut down cleanly.

However, as pointed out
bitcoin#33212 (review) and
bitcoin#33208 (comment) there
are other remaining cases where index state could be committed referring to
unflushed blocks, and this commit fixes those cases.

Add new interfaces::Chain::attachChain method and start moving index sync logic
out of src/index/ to src/node/. In this commit, move code that registers for
index validationinterface notifications. In later commits, move the actual sync
thread and add real blockConnected, blockDisconnected, and chainStateFlushed
implementations.

This commit does not change behavior in any way.

Split up and rename validationinterface callback functions
BaseIndex::BlockConnected to BaseIndex::IgnoreBlockConnected and
BaseIndex::ChainStateFlushed to BaseIndex IgnoreChainStateFlushed.

Most of the code in these functions was just ignoring spurious notifications,
so leave that code in place, but move the rest of the code to
BaseIndexNotifications::blockConnected and
BaseIndexNotifications::chainStateFlushed. In an upcoming commit the spurious
notifications will be dropped and the new BaseIndex::Ignore... methods will be
deleted entirely.

This commit does not change behavior in any way.

…BaseIndex destructor

This code never worked and was never used. Now that an m_handler member is
added, the incorrect usage is easier to detect and an assert can be added.

This commit does not change behavior in any way.

This commit does not change behavior in any way. Goal of this change is to
document behavior better and make it more explicit.

…ected

Start moving indexing code from BaseIndex::Sync() to blockConnected() so
indexes can be separated from the node and do not need to implement syncing.
More index-specific code will be moved from Sync() to blockConnected() in
upcoming commits and then the Sync method will be dropped.

This commit does not change behavior in any way.

…n handlers

This commit does not change behavior in any way.

…ion handler

This commit does not change behavior in any way.

…ed and Sync

This main thing this commit code does is make indexes start using
blockDisconnected() notifications instead of ignoring them and calling Rewind()
in blockConnected() after the fact. It also changes Sync() code to stop calling
Rewind() and instead call `blockDisconnected` in a loop. It moves most of the
code was that previously in Rewind() to blockDisconnected(), and some of it to
blockConnected(), which does not change behavior because Rewind was always
called before connecting a new block.

This commit tries to avoid changing behavior in any way. In particular, the
best block pointer is still not updated during the rewind, so indexes still
report status in the same way they did before this commit. (In the future it
would be good to consider updating best block pointer in a more consistent way,
but that is beyond the scope of this PR). Also failures during rewind still
cause fatal errors and interrupt the Sync() loop.

The only actual change in behavior should be that when blocks are disconnected,
the disconnects are processed earlier: as soon as blockDisconnected events
arrive instead of right before the next new block is connected. Additionally if
rewind occurs during an initial sync, it is now interruptible, and the rewind
loop will exit if `m_interrupt` is set.

Add BlockFilterIndex::m_filter_mutex to avoid TSAN errors in next commit moving
index sync thread from indexes to node. This mutex was previously neccessary,
but TSAN did not detect errors until the sync thread was moved.

No changes in behavior, this change is mostly move-only.

This commit does not change behavior in any way.

Add SyncChain method and improve Chain::attachChain behavior so it no longer
sends notifications from the sync thread at the same time as notifications from
validationinterface queue. This allows BaseIndex::m_ready variable and
surrounding logic to be dropped and indexing code to be simplified.

It is also a potential performance improvement since it avoids holding cs_main
while processing all indexing notifications. Previously cs_main was held while
sending some notifications in the attachChain sync thread.

In attachChain, avoid locking cs_main while calling prepare_sync method to run
index CustomInit code.

This is implemented by unconditionally spawning index sync threads, instead of
only spawning them conditionally when the index best block is not the chain
tip. An alternative implementation could avoid the cost of spawning the sync
thread by calling prepare_sync from the validationinterface notification
thread. But this would be slightly more complex, and also potentially slower if
any current or future index takes longer to initialize, since it would
initialize serially instead of in parallel with other index and node init code.

This commit does not change behavior in any way.

…ng code

This commit does not change behavior in any way.

Replace last m_blockman.UpdatePruneLock call with
interfaces::Chain::updatePruneLock call.