Fix ROCm EP build for gfx1152, add Docker build recipe

[zetyquickly]

Summary

• Fixes GPU_WARP_SIZE in onnxruntime/core/providers/rocm/cu_inc/common.cuh: constexpr int GPU_WARP_SIZE = warpSize; doesn't compile on this ROCm's bundled clang (23.0.0git) on either the host or device pass of this dual-compiled header — warpSize's host-visible form has a non-constexpr operator int(). Device pass now derives the value from __AMDGCN_CUMODE__ (0 = WGP/wave32 for RDNA chips like gfx1152, 1 = CU/wave64 for CDNA/GCN); host pass uses a fixed literal, since that value is never used for actual kernel execution there — genuinely dynamic runtime code already uses GPU_WARP_SIZE_HOST a few lines below.
• Adds a missing #include <cstdint> to optimizer_api.h (this GCC's libstdc++ dropped an implicit transitive include the file relied on).
• Adds docker/gfx1152/: a Dockerfile (ROCm 7.14 + full build toolchain on ubuntu:26.04) and a README with the full build recipe, why this is based on the v1.22.2 tag rather than main (ROCm EP was deleted from main in microsoft#25181 — no active development, Microsoft's stated direction is the MIGraphX EP instead), the two build flags needed to route around unrelated dead ends (onnxruntime_USE_COMPOSABLE_KERNEL=OFF, --disable_contrib_ops), and every environment gotcha hit along the way.

This branch has no realistic merge target upstream (main no longer has this code at all), so this base branch (v1.22.2-base, pointing at the v1.22.2 tag) exists purely so this PR's diff shows only the actual changes instead of the full divergence between v1.22.2 and main.

Test plan

• Built the wheel end-to-end inside the Docker recipe in docker/gfx1152/
• Installed the resulting wheel and ran Docling's ONNX-backed layout model (RT-DETR) through ROCMExecutionProvider on a real AMD Ryzen AI 7 PRO 350 (Radeon 860M, gfx1152) GPU — correct output, no crash

🤖 Generated with Claude Code