Filter Video Gen model dropdown to ltx2 runtime in a2v mode + auto-select largest model that fits

[atomantic]

Summary

• The Audio→Video mode previously showed every video model in the dropdown and rendered a warning telling the user to "switch the Model dropdown to one of the dgrauet entries" — which on installs where the dgrauet entries had been pruned from data/media-models.json referred to options that weren't in the list. The server enforces A2V_REQUIRES_LTX2 and would 400 on submit, so filtering client-side keeps the dropdown honest.
• <ModelSelect> for a2v now receives only runtime === 'ltx2' models (via a visibleModels memo). The model-validation useEffect now also checks mode-compatibility, so switching to a2v while an mlx_video model was selected auto-swaps to a compatible model on the next render. The previous bespoke auto-select inside handleModeChange was removed — one source of truth.
• For a2v, auto-select picks the highest-memory compatible model that fits within systemMemoryGb - 16 GB headroom (OS + text encoder + working set). On a 128 GB box this lands on ltx23_dgrauet_q8 (~25 GB); on a 32 GB box it lands on ltx23_dgrauet_q4 (~16 GB) — Q8 wouldn't fit the budget. If nothing fits, falls back to the smallest model so the user can still try and let the install banner / OOM surface the constraint.
• Server /api/video-gen/status now returns systemMemoryGb (rounded os.totalmem()) so the client can make the runnable-size decision. Additive, optional field — older clients ignore it; the new client gracefully falls back to POSITIVE_INFINITY budget when the field is absent.
• The inline warning under the audio uploader now only fires when visibleModels.length === 0 — i.e. no ltx2 models are installed at all — and the copy explains how to add a dgrauet entry to data/media-models.json and provision the runtime with INSTALL_LTX2=1 bash scripts/setup-image-video.sh.

Test plan

• Switch Video Gen mode to Audio. The Model dropdown only lists dgrauet (runtime: 'ltx2') entries.
• Switch back to Text / Image / FFLF / Extend. Full model list reappears.
• On a machine with both ltx23_dgrauet_q4 and ltx23_dgrauet_q8 installed and ≥ 48 GB RAM, switching into a2v auto-selects Q8 (highest that fits).
• On a machine with both installed but < ~41 GB RAM (16 GB reserve + 25 GB model), auto-select lands on Q4.
• When data/media-models.json contains no runtime: 'ltx2' entries, switching to a2v shows the new "no a2v-compatible models installed" warning and the dropdown is empty.
• Connect a new client to an older server that doesn't return systemMemoryGb — auto-select still picks the largest dgrauet model (infinite budget fallback).
• GET /api/video-gen/status includes systemMemoryGb: <integer> matching os.totalmem() rounded to GB.