docs: add Model Backends page and document AMD ROCm support

[zetyquickly]

Summary

• Adds docs/usage/model_backends.md, organizing Docling's models along the runtime-backend axis (Pure PyTorch, ONNX, MLX, vLLM, other), wired into the nav and cross-linked from the model catalog.
• Adds an AMD ROCm section to gpu.md: device="cuda" is correct on a ROCm-built PyTorch too (it aliases HIP under torch.cuda), with install instructions for the standard ROCm wheel index and AMD's multi-arch per-gfx index for newer APUs, how to verify a wheel actually ships kernels for your GPU's gfx target before trusting it (mismatches segfault/hang rather than raising cleanly), and known caveats (ONNX has no ROCm execution provider yet, cuda_use_flash_attention2 assumes the CUDA-only flash-attn wheel, Nemotron OCR and WhisperS2T are CUDA/CTranslate2-only).
• Verified end-to-end on an AMD Ryzen AI 7 PRO 350 (Radeon 860M, gfx1152, ROCm 7.14): the default layout model and TableFormer ran their real GPU kernels through the standard PDF pipeline and produced correct output (text extraction + a structured table), with zero Docling code changes.

Test plan

• Manually verified torch.cuda.get_arch_list() includes gfx1152 after installing from repo.amd.com/rocm/whl-multi-arch
• Ran DocumentConverter with AcceleratorOptions(device=AcceleratorDevice.CUDA) against a real PDF on the AMD iGPU; ConversionStatus.SUCCESS, text + 1 table extracted correctly
• Docs-only change; no code paths modified

🤖 Generated with Claude Code

[github-actions]

❌ DCO Check Failed

Hi @zetyquickly, your pull request has failed the Developer Certificate of Origin (DCO) check.

This repository supports remediation commits, so you can fix this without rewriting history — but you must follow the required message format.

---

🛠 Quick Fix: Add a remediation commit

Run this command:

git commit --allow-empty -s -m "DCO Remediation Commit for zetyquickly <zetyquickly@googlemail.com>

I, zetyquickly <zetyquickly@googlemail.com>, hereby add my Signed-off-by to this commit: bcddf8cb2009d4cf1b1e1f633fecdb679063f532"
git push

---

🔧 Advanced: Sign off each commit directly

For the latest commit:

git commit --amend --signoff
git push --force-with-lease

For multiple commits:

git rebase --signoff origin/main
git push --force-with-lease

More info: DCO check report