fix: hide score on submit + total denominator + projector chart cleanup

Three small UX/fairness tweaks from manual live testing:

1. Post-submit "wait for reveal" screen: show only the response time, no
   score. The +score reveal leaked correctness — any positive number =
   correct, zero = wrong — short-circuiting the "stop and think" beat
   the reveal pause was supposed to enforce. Time stays as the
   engagement signal; score now waits for the instructor reveal.

2. Final-screen "Correct X / Y" denominator is now total_questions
   instead of questions_answered. Missed questions are scored zero, so
   they belong in the denominator visibly. Server adds total_questions
   to the session_ended payload.

3. Projector score-distribution: drop the in-chart count labels (they
   collided with each other and with the median tag at small N), restore
   the previously-computed-but-not-rendered x-axis tick labels at the
   bottom. Stats line at the foot keeps n / mean / max.

Also: short-circuit the per-submit instructor + presence broadcasts
when no instructor / projector is connected (no listener, no DB work).
The 50-student load test was tight on margin against its 2 s
time_limit; with the new presence_message / live_histogram_message DB
queries firing on every submit, the margin disappeared on busy boxes.
Conftest fixture also bumped to 8 s per question for the same reason —
gives breathing room for sequential WS submits in the load test.

71/71 pytest green.

static/style.css

@@ -1155,6 +1155,16 @@ h2.question-text.small {
   letter-spacing: -0.04em;
   line-height: 1;
 }
+/* Unit suffix (e.g. "s" after a duration) — small, muted, baseline-sized
+ * so it reads as a tag, not part of the number. */
+.big-score .unit {
+  font-size: 0.32em;
+  color: var(--muted);
+  letter-spacing: 0;
+  margin-left: 6px;
+  vertical-align: 0.55em;
+  font-weight: 500;
+}
 
 .spinner {
   width: 22px;