fix(dist): plug Remove silent-swallow and hint-replay abandon on transport errors

[hyp3rd]

Two symmetric audit fixes for the distributed memory backend:

Remove forward promotion (removeImpl):
removeImpl previously blackholed the ForwardRemove result with _ = transport.ForwardRemove(...), so a Remove against a downed primary silently succeeded while the stale value lingered on every owner. The new forwardOrPromoteRemove helper mirrors the handleForwardPrimary promotion contract: on any non-nil transport error, if the local node is a replica owner it applies the remove locally and fans out to peer replicas via applyRemove(replicate=true); otherwise it returns the error. Promotions bump the shared dist.write.forward_promotion counter so Remove and Set promotions are observable on the same instrument.

Hint replay retention (processHint):
processHint previously retained a hint only when the error matched the in-process ErrBackendNotFound sentinel. Production HTTP/gRPC transports surface net.OpError, io.EOF, and
context.DeadlineExceeded for a briefly-unreachable peer, causing the hint to be abandoned on its very first replay attempt instead of being retained through the outage (recovery on :8083 timed out after 60s). The hint TTL (WithDistHintTTL) still bounds total retry time, so a permanently-broken target still drains. The deprecated hintedDropped / migrationHintDropped OTel counters remain registered for stability but now only bump on queue-capacity overflow.

Tests:

• Add TestDistRemove_PromotesOnGenericForwardError: chaos at DropRate=1.0, asserts promoted Remove returns nil, clears local copy, bumps counter.
• Add TestDistHintReplay_RetainsOnGenericReplayError: 150ms chaos window, heals, asserts hint replays onto recovered peer.
• Rename TestMigrationHint_TransportErrorBumpsDroppedCounter -> TestMigrationHint_TransportErrorKeepsEntry to pin the new contract.