smp-server: fix notification store key leak

[shumvgolove]

The notification store kept a key for every notifier that ever received a notification: expiry trimmed the list but not the key, and delivery emptied the list but not the key. deliverNtfsThread sends every key to getQueueNtfServices (one notifier_id IN ? query, read in full) every 1.5 s, so the cost of each tick grew with every notifier notified since restart.

• Expiry removes a key when its list becomes empty, and storeNtf is atomic, so a notification cannot be stored to a removed key's TVar.
• After each delivery tick, keys whose lists were emptied are removed (deleteEmptyNtfs), each in its own small transaction, only if the list is still empty.

Delivery loop with no traffic (ntfloop):

Keys	Allocation	CPU	Peak memory in use
0	0 MiB/s	0	280 MiB
100k	168 MiB/s	0.42 cores	530 MiB
300k	263 MiB/s	0.76 cores	913 MiB

20k notifications delivered to a subscribed ntf client, then idle (ntfdeliver):

	Before	After
keys kept	20,000	0
idle allocation	36 MiB/s	0.2 MiB/s
idle CPU	0.18 cores	0.01 cores

New tests in MsgStoreTests (notification store): expired and empty keys are removed; only flushed keys that are still empty are removed.