FIX: count languages in the Multilingual run size estimate

[Utkarsh Bahuguna (u7k4rs6)]

Fixes #3083.

Multilingual now has its own _estimate_run_size_async, shaped like the run it builds:

• Baseline: one unit per seed group (when enabled)
• Translation: seed groups x techniques x languages
• Random translation: seed groups x techniques

The language count comes from a new _get_language_selection(), which holds the validation _resolve_languages_async already had (both selectors, empty list, num_languages range). The run path uses the same helper, so explicit languages get deduped the same way in both. Resume still replays the stored set.

Same repro on this branch:

default (5 random languages): estimate total=35 | initialized units=35
languages=[8 languages]:      estimate total=50 | initialized units=50

Test: estimate equals the initialized unit count for the default run, num_languages, duplicate languages, each strategy alone, baseline off and two techniques. 6 of the 7 fail on main (random-translation-only happens to match). tests/unit/scenario and backend: 3820 passed.