voteTracker held every accepted vote in Voters and again in
Counts[proposal].Votes. Since the previous release generates bundles
from Voters, the copies have no reader: stop maintaining them. Each
vote is about 780 bytes encoded and the copies went into every crash
recovery snapshot, about a third of a mainnet cert-step snapshot.

The field stays, omitted when empty, so that snapshots written by the
previous release still decode; it restores them with the copies, which
it does not read. Releases before the previous one would restore a
snapshot without the copies with incomplete bundles, so this change
ships one release after the bundle change.

genBundle built a threshold's bundle from Counts[proposal].Votes, a
second copy of the votes already held in Voters. Build it from Voters,
so that a persisted state without the per-proposal copies restores
with complete bundles; the next release stops persisting them. The
copies are still maintained here, so that a state persisted by this
release restores on the previous one.

Validate a restored vote tracker before using it: the counts must be
the weight of its voters, the equivocators count the weight of its
equivocators, and no sender in both maps. A state that fails is not
restored; the node starts the round fresh, as it does for a state it
cannot decode.

Non breaking changes to ARC60 to support PQ schemes. CAIP was also
removed and replaced with a security section noting that address auth is
non-trivial on Algorand due to rekeying. A future ARC that leverages
ARC60 can standardize SIWA

The persisted state is mostly the votes of the round's steps, and it
holds their content several times over: every vote carries the proposal
value, the bundle that proved a threshold repeats the signatures of the
votes that formed it, and msgpack writes every field name. zstd at
level 3 replaces the repeats with references and leaves the signatures
and proofs: a mainnet snapshot shrinks from 433 KB to 105 KB, for
0.75 ms to encode and 0.25 ms to decode. The snapshot is written twice
per round, so this removes most of the crash database's disk writes.

decode recognizes a zstd frame by its magic number and still reads the
uncompressed snapshots of earlier releases.

Add MakeLongWALAccessor, whose connections checkpoint the write-ahead
log every 8192 frames instead of SQLite's 1000.

The net result is checkpoint happens less frequently reducing number of pages written.

At the current mainnet state it is reduction from 22% WAL pages down to 3% copied.