docs: add gfx1200 (Navi 44) alongside gfx1201 for RDNA4 support

[0xDELUXA]

Motivation

The README News entry and the two RDNA4 recipes added in #811 name only gfx1201 (Navi 48 - RX 9070 / RX 9070 XT / AI PRO R9700). Its sibling RDNA4 chip gfx1200 (Navi 44 - RX 9060 / RX 9060 XT) rides the exact same code path but is never mentioned, so a 9060 / 9060 XT owner has no signal that ATOM covers their card. This PR names gfx1200 alongside gfx1201 everywhere the arch is declared or built, so the docs match the support that already exists.

Technical Details

This is intentionally docs-only, mirroring how gfx1201 itself shipped.

#811 deliberately removed every arch-specific dispatch from ATOM Python (the _is_gfx1201 / _detect_gfx1201 gates, the ATOM_NATIVE_TRITON_ATTN env var, the selector.py allowlists, the scripts/gfx1201/ config workaround) and pushed the fallbacks down into aiter, leaving ATOM 100% capability-driven. There is therefore no arch gate in ATOM to extend for a new RDNA4 chip - the only place the arch is named is the documentation.

The kernel-level RDNA4 support gfx1201 relies on already covers gfx1200. The aiter PRs #811 depends on were written for the RDNA4 family and name gfx1200 explicitly alongside gfx1201:

• [opus] gfx1201/gfx1200 (RDNA4) support: parse + buffer rsrc + WMMA w32/w64 aiter#3236 - gfx1201/gfx1200 RDNA4 support: parse + buffer rsrc + WMMA w32/w64
• [dtypes] add gfx1200 / gfx1201 (RDNA4) to FP8 dtype map aiter#3332 - dtypes.fp8 → float8_e4m3fn on gfx1200/1201

So the only per-chip difference is the GPU_ARCHS string when building aiter (ROCm/aiter#3846) and typically less VRAM.

Changes:

• README.md - News entry now reads Navi 4 (RDNA4 / gfx1200, gfx1201) and lists the RX 9060 / RX 9060 XT (Navi 44) cards.
• recipes/Ministral-3-8B.md - build-aiter prerequisite note covers GPU_ARCHS=gfx1200 as well as gfx1201.
• recipes/Qwen3-8B-FP8.md - same for the inline prerequisite, plus the rmsnorm_quant JIT note.

Test Plan

• Confirmed via git grep that no gfx1201 literal exists in ATOM Python/Rust on main (only gfx1250 / gfx94x / gfx95x are special-cased), so this change cannot touch a code path - it is documentation only.
• Traced the gfx1201 enablement to [gfx1201] Mistral-3 + Qwen3-8B-FP8 on RDNA4 via native triton attention #811 and confirmed its final merged diff adds no arch dispatch, and that its aiter dependencies (ROCm/aiter#3236, ROCm/aiter#3332) already name gfx1200.
• Verified the patch applies cleanly to main and that the rendered Markdown links/anchors are unchanged.

Test Result

Docs-only change, so no code paths are exercised. atom/aiter has no Windows support and I'm a Windows-only user, so I can't run this end-to-end locally. The benchmark/accuracy tables are therefore left labeled gfx1201-measured and no numbers are claimed for gfx1200 - the change is scoped to "same path, same build prerequisite." Happy for a maintainer with gfx1200 hardware to verify, and I'll add a results block if it reproduces.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[0xDELUXA]

Unfortunately, I can’t properly test gfx1200 support locally since I don’t use Linux, and aiter doesn’t have Windows support: ROCm/aiter#2869.