v0.2.1 - Model download with %% progress

Backend:
- download.py module: drives ./hf-download.sh <repo> [-c --copy-parallel] over SSH, parses tqdm output (regex matches '8%|...| 2.06G/25.1G [03:20<18:35, 20.6MB/s]') into percent + bytes done/total + elapsed + ETA + rate
- DownloadManager: in-memory job tracking with asyncio.Lock (one download at a time)
- POST /api/download, GET /api/download/{id}, SSE /api/download/{id}/stream
- Phase detection: Connecting / Fetching N files / Downloading / Copying to peer Sparks / Done

Frontend:
- '+ Download a new model' button next to LLM swap section title
- Inline form: HF repo text field + solo/cluster radio + Cancel/Start
- Progress UI: spinner, elapsed timer, phase label, percent fill, stats line (bytes/rate/ETA), collapsible raw logs

Package: bump 0.2.1:0

package/startos/versions/v0_1_0.ts

@@ -1,10 +1,10 @@
 import { VersionInfo, IMPOSSIBLE } from '@start9labs/start-sdk'
 
 export const v0_1_0 = VersionInfo.of({
-  version: '0.2.0:0',
+  version: '0.2.1:0',
   releaseNotes: {
     en_US:
-      'Always-on services panel: dashboard now has cards for Parakeet and Magpie with Start/Stop/Restart buttons and live container state (uptime, restart count). Configure Sparks adds optional per-service host/container fields so Parakeet or Magpie can live on Spark 1 (or anywhere) instead of being hard-wired to Spark 2.',
+      'Model download from the dashboard. Click "+ Download a new model", paste a Hugging Face repo (e.g. RedHatAI/...), choose solo or cluster, and watch %% progress with bytes, rate, and ETA. Drives ./hf-download.sh on Spark 1 over SSH and parses the tqdm output.',
   },
   migrations: {
     up: async ({ effects }) => {},