#1694 mcp + docs: let the swift family take IQ3_S too (follow-up to #1688)

Description

Follow-up to #1688: `setup.py` and the MCP server keep **two copies** of the model table. #1688 drops the `families=("qwen",)` gate in `setup.py`; this PR drops it in the MCP server's `FALLBACK_MODELS` too, so `strata_install {family: swift, model: IQ3_S}` through the MCP tools stops being rejected, and fixes the three doc lines that still say IQ3_S is original-model-only.

## What #1688 does not cover

- `tools/strata_mcp.py` — `FALLBACK_MODELS["IQ3_S"]` carries its own `families: ("qwen",)`. After #1688 merges, the setup path accepts `swift + IQ3_S` while the MCP install path (docs/AI_SETUP.md's flow for AI assistants) still rejects it.
- `tools/test_strata_mcp.py` — the validation test asserts `swift + IQ3_S` is rejected ("has no IQ3_S"); flipped to assert it is accepted and the plan's setup command carries `--family swift --model IQ3_S`.
- `docs/MODELS.md` / `docs/DETAILS.md` — three lines: the IQ3_S table row "(original model only)", the Swift paragraph "(no IQ3_S)", and the DETAILS sentence "is only published for the original model, not for Swift 1.5".

## What was measured

The part #1688's test plan lists as not tested — an actual download/run of the tier — is running here:

RTX 3090 pair, engine 0.1.40.3, Swift 1.5 IQ3_S, 200K context, `--spec 4 --mtp`, `--kv int8 --kv-resident 32768`:
- decode 91-137 tok/s
- decode expert cache hit rate 96.5-99.0%
- KV streaming: 98.6-99.6% of block reads hit VRAM
- draft acceptance 80-95% on most verify windows

## Test plan

- `tools/test_strata_mcp.py` — 24 tests, all pass (with the flipped swift+IQ3_S assertion).
- `tools/test_setup_choices.py`, `test_setup_config.py`, `test_setup_configs.py` — all pass (untouched by this PR; run as a cross-check).

No setup.py changes here on purpose, so this merges cleanly on top of #1688 (or before it — the files do not overlap).