Replace Khepri topic routing projection with trie + ordered_set (v4)

[ansd]

Resolves #15588.

The previous Khepri topic routing projection (v3) stored topic bindings as sets:set(#binding{}) inside trie leaf nodes. This design had a major performance drawback:

On the insertion/deletion path (in the single Khepri Ra process), every binding change required a read-modify-write of the entire sets:set(), making it O(N) in the number of bindings at that leaf. With many MQTT clients connecting concurrently (each subscribing to the same topic filter), this made the Ra process a bottleneck.

Another less severe performance issue was that the entire binding was being copied including the binding arguments containing the MQTT 5.0 subscription options such as:

{<<"x-mqtt-subscription-opts">>,table,
 [{<<"id">>,unsignedint,1},
  {<<"no-local">>,bool,false},
  {<<"qos">>,unsignedbyte,0},
  {<<"retain-as-published">>,bool,false},
  {<<"retain-handling">>,unsignedbyte,0}]}]

Replace the single ETS projection table with two purpose-built tables:

1. Trie edges table (ETS set, read_concurrency=true):

   • Row: {{XSrc, ParentNodeId, Word}, ChildNodeId, ChildCount}
   • XSrc = {VHost, ExchangeName} (compact 2-tuple of binaries)
   • NodeId = root | reference()
   • ChildCount tracks outgoing edges for garbage collection

2. Leaf bindings table (ETS ordered_set, read_concurrency=true):

   • Key: {NodeId, BindingKey, Dest}
   • Stored as 1-tuples: {{NodeId, BindingKey, Dest}}
   • No value column; all data is in the key to minimize copying

The trie structure preserves O(depth * 3) routing complexity regardless of the number of overall bindings or wildcard filters. At each trie level, we probe at most 3 edges (literal word, <<"*">>, <<"#">>), each via ets:lookup_element/4 which copies only the ChildNodeId (a reference).

The ordered_set for bindings provides:

• O(log N) insert and delete per binding (no read-modify-write)
• The binding key (needed for MQTT subscription identifiers and topic aliases) is part of the key, so it is returned directly during destination collection without additional lookups

Collecting destinations at a matched trie leaf uses a hybrid strategy:

• Fanout 0-2 (the common case: unicast, device + stream): up to 3 ets:next/2 probes. Each ets:next/2 call costs O(log N) because the CATree (used with read_concurrency) allocates a fresh tree traversal stack on each call.
• Fanout > 2: ets:select/2 with a partially bound key does an O(log N) seek followed by an O(F) range scan. The match spec compilation overhead amortises over the larger result set.

ets:lookup_element/4 (OTP 26+) returns a default value on miss instead of throwing badarg, and copies only the requested element on hit. This avoids both exception overhead (misses are common during trie traversal of <<"*">> and <<"#">> branches) and unnecessary data copying (we only need the ChildNodeId, not the full row).

Trie node IDs are ephemeral (the tables are rebuilt when the Khepri projection is re-registered). make_ref() is fast, globally unique within a node, and has good hash distribution for the ETS set table.

When a binding is deleted, the trie path from root to leaf is collected in a single downward walk (trie_follow_down_get_path). Empty nodes are then pruned bottom-up: a node is empty when its ChildCount is 0 and it has no bindings in the ordered_set table.

Benchmarks below were run with 500K routing operations per scenario (on the same machine, back-to-back between main (v3) and this commit.

Significant insert/delete improvements:

Churn insert (8K bindings, 4 filters/client): ~1,120 vs ~810 ops/s (+38%)
v3 did a read-modify-write of sets:set() per binding; v4 does
a single ets:insert into the ordered_set plus trie edge updates.

MQTT device insert (20K bindings): ~650 vs ~420 ops/s (+55%)
Same mechanism as churn insert. Particularly impactful when many
clients share the same wildcard filter (e.g. "broadcast.#"),
since v3's sets:set() grew with each client while v4 inserts
are O(log N) regardless.

Same-key fanout insert (10K): ~415 vs ~290 ops/s (+43%)
The worst case for v3: all 10K bindings share the same key,
so each insert copies and rebuilds the growing sets:set().

Routing improvements:

MQTT unicast (10K devices, 20K bindings): ~460K vs ~250K ops/s (+80%)
Each route matches 1 queue among 10K unique exact keys plus
10K queues sharing "broadcast.#". v3 stored bindings in the same
ETS row as the trie edge, so every trie lookup copied the entire
sets:set(). v4 separates trie edges (small rows, set table) from
bindings (ordered_set), so the trie walk copies only references.

Large fanout (10K queues, same key): ~3,100 vs ~1,170 ops/s (+165%)
v3 copied a 10K-element sets:set() out of ETS in a single
ets:lookup, then called sets:to_list/1. v4 uses ets:select/2
with a partially bound key, which does an O(log N) seek and
then an efficient O(F) range scan without intermediate set
conversion.

MQTT broadcast (10K fanout): ~0.6 vs ~0.9 ms/route (+50%)
Same mechanism as above.

Scenarios with no significant change (within benchmark noise):

Exact match, wildcard *, wildcard #, mixed wildcards, and many
wildcard filters showed no clear difference. Both v3 and v4 use
a trie walk, so routing speed is comparable when the fanout is
small and the bottleneck is trie traversal rather than destination
collection.

[ansd]

I mark this PR as Ready for review but we should not merge this PR until all of the following are done:

• More benchmarks checking for any major regressions (e.g. via MQTT or AMQP or definition import workloads) @mkuratczyk couldn't spot regressions. @the-mikedavis in case you have still have your benchmark tools available from your previous topic exchange work, could you please also run them against this PR?
• User feedback
• Khepri changes discussed in Replace Khepri topic routing projection with trie + ordered_set (v4) #15619 (comment)