[Backport] Implement accurate UTXO cache size accounting

[furszy]

This adds a basic foundation for accurate memory counting (memusage.h, with some implementation-specific assumptions, but at least the tree node and hashtable node implementations used are very generic and straightforward, so likely accurate for several systems). On the tested system at least, they are exact, ignoring memory fragmentation (tested using a single binary creating and modifying large amounts of different configurations of these data structures, and observing total resident set size afterwards). I expect this to be useful for other resource-limiting subsystems later on.

Then, this is used to implement accurate memory usage counting for CCoins objects, and efficiently computed memory usage counting for CCoinsViewCache (using cached totals, and increments/decrements on updates through CCoinsModifier). The existing CCoinsViewCache randomized simultation unit test is extended to also verify the correctness of the cached memory usage totals. Changing any of the cached total update statements in coins.cpp breaks the unit test.

Finally, the internal flush triggering mechanism is changed to use the memory usage mechanism rather than transaction count of pcoinsTip.

Coming from upstream@6102 .

--- Needs more testing ---

[random-zebra]

Pretty cool. Concept and code ACK.
Will do some testing.

[random-zebra]

Tested resyncing testnet with default (100 MiB) dbcache, which reserves 57.6 MiB for nCoinCacheUsage, resulted in 16 flushes to disk.

Resynced again, this time setting -dbcache=1700 (17 times the default), which translates in a 1107 MiB nCoinCacheUsage, did not result in a single database flush as expected (final pcoinsTip's cache memory usage at testnet block 1579784 is 764 MiB).

NOTE: In order to do this test, and flush only when needed (i.e. only when fCacheLarge), had to apply this patch:

Index: src/main.cpp
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
--- src/main.cpp	(revision 96b7ec0f6986ad609822d35bf24cfff643ac73c3)
+++ src/main.cpp	(date 1589465455694)
@@ -2741,7 +2741,7 @@
     }
     LogPrint(BCLog::BENCH, "- Disconnect block: %.2fms\n", (GetTimeMicros() - nStart) * 0.001);
     // Write the chain state to disk, if necessary.
-    if (!FlushStateToDisk(state, FLUSH_STATE_ALWAYS))
+    if (!FlushStateToDisk(state, FLUSH_STATE_IF_NEEDED))
         return false;
     // Resurrect mempool transactions from the disconnected block.
     for (const CTransaction& tx : block.vtx) {
@@ -2814,11 +2814,8 @@
     nTimeFlush += nTime4 - nTime3;
     LogPrint(BCLog::BENCH, "  - Flush: %.2fms [%.2fs]\n", (nTime4 - nTime3) * 0.001, nTimeFlush * 0.000001);
 
-    // Write the chain state to disk, if necessary. Always write to disk if this is the first of a new file.
-    FlushStateMode flushMode = FLUSH_STATE_IF_NEEDED;
-    if (pindexNew->pprev && (pindexNew->GetBlockPos().nFile != pindexNew->pprev->GetBlockPos().nFile))
-        flushMode = FLUSH_STATE_ALWAYS;
-    if (!FlushStateToDisk(state, flushMode))
+    // Write the chain state to disk, if necessary.
+    if (!FlushStateToDisk(state, FLUSH_STATE_IF_NEEDED))
         return false;
     int64_t nTime5 = GetTimeMicros();
     nTimeChainState += nTime5 - nTime4;

Not sure why we are forcing FLUSH_STATE_ALWAYS in ConnectTip (if the first of a new file: coming from #119), and in DisconnectTip. But it can be discussed/addressed in a separate PR.

[furszy]

Awesome feedback and discovery @random-zebra 👌 .

Just updated the PR with the typo correction.

[random-zebra]

ACK 9880245

[Fuzzbawls]

ACK 9880245