Add descriptions to paper posts and fix factual errors

[bits-bytes-nn]

All 28 paper posts now declare description: (120-158 characters; starts with the method name; numbers only from the post's TL;DR or abstract). Before, _plugins/post_description.rb derived each from the introduction, and 18 of 28 opened on "대규모 언어 모델(…)의 발전…".

Hand fixes to generated posts (source value vs. what the post said):

• Mistral 7B, lines 18 and 32: 70억 parameters (7B); post said 7억.
• DeepSeek LLM, TL;DR: 0.5243 and 0.4757 are the exponents of compute in the optimal model and data scale (M_opt ∝ C^0.5243, D_opt ∝ C^0.4757); post said 52.43% / 47.57% of the budget. The model that beats GPT-3.5 is DeepSeek LLM 67B Chat, not the base 67B.
• RAPTOR, TL;DR: Recursive Abstractive Processing for Tree-Organized Retrieval; post said "Recursive Abstraction Processing Through Tree-based Organization and Retrieval".
• Gemma 2, TL;DR: 2B and 9B are distilled from a larger teacher (the 7B teacher is only the ablation); distillation lifts the 2B average by 7.4 points (60.3 to 67.7); post said a 7B teacher and 7.4%.
• Llama 3, TL;DR: 405B instruct scores MMLU 87.3% and GSM8K 96.8% (Table 2); post said 85.2% (the pre-trained score) and 89.0% (the HumanEval figure).
• Qwen3, two places: models from 6억 (0.6B) parameters; post said 0.6억.
• Zep, TL;DR and §4.3.2: LongMemEval gains of 15.2% (55.4% to 63.8%, gpt-4o-mini) and 18.5% (60.2% to 71.2%, gpt-4o) are relative; post said percentage points.

Not changed, unverified: Gemini 2.5 line 30 says the model was trained with a "Constitutional AI 접근법"; the report may not describe it that way.

Gates: tests, post lint, strict build, htmlproofer, validate-site (description uniqueness).