#2557 docs: the 64 GB row recommends IQ3_S, as setup does since 0.1.44 (README in 7 languages, MODELS.md, AI_SETUP.md)

Descrição

**Why.** f3ee1819 made IQ3_S setup's recommended size:

> setup: IQ3_S (Qwen, Swift) and UD-Q4_K_XL / UD-IQ4_XS (Unsloth) are the recommended sizes and the defaults where the RAM fits them; the largest that fits otherwise.

and `setup.py` says `RECOMMENDED_SIZES = {"qwen": ("IQ3_S",), "swift": ("IQ3_S",), ...}   # owner's pick, 0.1.44`. That commit touched no docs, so the README's model table and MODELS.md / AI_SETUP.md still told a 64 GB user that IQ2_XS is the recommended size, while the installer's menu marks IQ3_S "(recommended)" and `--yes` installs it (`ram_gb` 62 for IQ3_S; `ram_gb()` reads the total RAM in GiB). This PR follows that choice in the docs; it does not argue for either size.

**What it changes** (one change, in every file that states it; 12 lines):
- README.md and its six translations: the 64 GB row is now "**IQ3_S** (recommended), or IQ2_XS / IQ3_XXS (faster)"; the "why" cell is unchanged ("IQ3_S is the best and the slowest").
- docs/MODELS.md: the same row; "(recommended)" moves from IQ2_XS to IQ3_S in the size table; "Not sure?" now says what setup pre-selects: IQ3_S from 62 GB of RAM, else the largest size the RAM fits (IQ2_XS at 48 GB; IQ3_XXS between 60 and 62 GB, as `recommended_size()` does).
- docs/AI_SETUP.md: the 64 GB row is `--family qwen --model IQ3_S` (recommended), IQ2_XS / IQ3_XXS faster.
- docs/INSTALL.md (Docker): no longer calls the image's default `MODEL=IQ2_XS` "the recommended one"; it says IQ2_XS fits 48 GB and setup recommends IQ3_S from 62 GB.

**What it leaves alone.**
- The 32, 48 and 96 GB rows (they already match setup), every number, and all code.
- **Docker's default** (`docker-entrypoint.sh` `MODEL="${MODEL:-IQ2_XS}"`): unchanged. Open question for you: should the image follow setup (IQ3_S) too? It runs on any RAM size, so IQ2_XS may be deliberate there.
- docs/DETAILS.md's "Which model?" section, which still lists three sizes and calls IQ3_XXS the best; that is a separate fix.

**Note.** docs/AI_SETUP.md line 75 here sits next to line 76 in #2550 (Swift sizes, Unsloth on AMD); whichever merges second needs a one-line rebase, which I will do.

**How it was checked.** Against `setup.py` (`RECOMMENDED_SIZES`, `recommended_size()`, `MODELS[...]["ram_gb"]`, `ram_gb()`), and by a sub-agent reviewer for both that and the six translations (faithful, plural agreement, same structure as the neighbouring rows).

🤖 Generated with [Claude Code](https://claude.com/claude-code)