Clean dataset: remove noise columns, strip ads/boilerplate, propagate prompts to same-batch images

Cleaned telegram-channel-dataset for VLM LoRA training.

Changes:
- Removed 18 noise columns (Telegram metadata, author info, file paths)
- Kept 6 essential columns: message_id, datetime, width, height, image, text
- Removed 741 rows (orphan images without prompts, ads, promos, channel announcements)
- Propagated prompts to same-batch image rows (same-generation images share the prompt text)
- Stripped channel boilerplate: bot instructions, hashtags, source credits, model name prefixes, emoji title lines, channel footer ads
- 1,050 rows remain with clean structured Chinese/English prompts

dataset.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c821a248da33469ed15377dff57f767f4772cb1026543460c08b8bf83d725d24
-size 399713823
+oid sha256:5363065ac5e7b5ab4e675db7bd875563ea592fb6b981daeabb74c8a5b8b71f09
+size 237147801