Fixes flaky remove_listener subscriber test

[johnnyt]

Why

test/statifier_ui/trace/subscriber_test.exs - "remove_listener/2 stops
delivery to that pid" - failed intermittently: on 2026-08-22 it broke one
full-gate run on a branch that never touched the file, passed on immediate
rerun, and stayed green across three consecutive later gates. A timing race,
not branch-related (sui-hmm).

What

Test-only change. The test called add_listener/2 (a call from the test
process) immediately after start_early!, racing the initialize burst's
{:statifier, ...} messages (sends from the session process) into the
subscriber's mailbox - cross-sender delivery order is not guaranteed, so the
listener occasionally caught burst messages that later tripped
refute_receive.

The test now registers the listener at start_link time via
start_early!(..., listeners: [self()]), so it deterministically sees the
whole burst; it asserts seq 0..5 arrive (delivery was live), removes the
listener with the synchronous call before driving the event, and ends on
refute_received - the subscriber's final stats reply is ordered after any
fan-out send, so no sleep or timeout is needed.

Notes

• No Subscriber semantic change: fan-out and add/remove_listener behavior are
  untouched, so the just-merged inspector assembly's attach/catch-up usage is
  unaffected. Full suite green on the rebased tree (3b43827), including the
  new catch-up tests.
• Evidence: 100 runs of the test file under 8 busy-loop processes of CPU
  load, before vs after - before: 7/100 failures, all this test; after:
  0/100. (One further before-run iteration failed differently because the
  checkout was pulled mid-run - a torn-state artifact, not a flake.)
• Gate: full gate green - 427/427 tests, 92.3% coverage, Dialyzer clean;
  Gettext and Sobelow are declared not applicable in this repo.

Closes sui-hmm.