Reconcile remote registry snapshots with a map instead of an ordset

[vheathen]

reconcile_remote_registry_snapshot/3 builds the snapshot names with ordsets:from_list/1 and calls ordsets:is_element/2 once per entry the node already holds for that remote node, so the reconciliation is quadratic.

After an ordinary join or reconnection the receiver holds no entries for that node, because they were purged when the remote scope went down, so the cost stays hidden. It shows when an ack_sync arrives from a node whose entries are already in the table. One way to get there: init/1 calls net_kernel:monitor_nodes(true) before handle_continue(after_init) sends discover to nodes(), so a node that connects in between gets discover twice and answers with two full snapshots. The second one is compared against everything the first one inserted, inside the scope process, and every register on that node waits behind it.

Membership test alone, N snapshot names against N existing entries, OTP 27.3.4, median of three runs:

entries	ordsets	map
10 000	748 ms	1 ms
30 000	7 185 ms	5 ms
100 000	96 424 ms	20 ms

The fix uses a map and is_map_key/2. The map is built with maps:from_keys/2 from OTP 24 and with maps:from_list/1 below it. The suite cannot cover that fallback, because it needs peer and so OTP 25, so I checked it by hand on OTP 23: src compiles and the fallback is the branch compiled in.

The new test case fails on the current code with 50 308 359 reductions for a 10 000 entry snapshot and passes with the fix at about 281 000. Tests pass on OTP 25 through 29 and dialyzer is clean.

[vheathen]

@ostinelli may I ask you to look if this PR makes sense for you? We are switching to syn, and in some scenarios it helps quite much.