ILGPU v2.0: Docker Support + AI Readiness

[m4rs-mt]

This PR packages everything needed to run the compiler service in a container, deploy it to a remote host, and redesigns the canonical CI workflow to use those images for GPU-bound jobs. It also introduces CLAUDE.md — an onboarding file that helps AI coding assistants navigate the repository.

This PR is deliberately narrow — it contains only infrastructure / tooling / process files. No source-code changes.

The first three commits are PR placeholders for the runtime-library, AOT-compiler, and compiler-bindings PRs; once those land on master they collapse into real merge commits.

A note on authorship - AI warning

Much of the test-infrastructure work in this PR was done with AI-assistant pair programming (Claude via Claude Code).

What's in this PR

Docker images for the compiler service (Src/docker/, 5 files)

Three Dockerfiles, one per GPU-bound backend:

• Dockerfile.cuda — NVIDIA CUDA toolkit + ILGPUC.CompilerService for emitting PTX/cubin.
• Dockerfile.rocm — AMD ROCm + hipcc for HIP code generation.
• Dockerfile.opencl — Intel ocloc + clang-for-OpenCL for OpenCL backends.

Plus:

• Src/docker/run.sh — multi-image launcher that selects the right image based on the target backend requested.
• Src/docker/README.md — explains how to build, tag, and publish the images and how they're consumed from CI.

Deployment scripts (Src/scripts/, 2 files)

• deploy-compiler-service.sh — systemd-based remote deployment helper. SSHes to a target host, pulls the relevant docker image, and registers / starts a systemd unit pointing at it.
• README.md — documents the deploy flow, required host prerequisites, and how to roll back a deployment.

Root configuration

• .dockerignore — excludes test artefacts, IDE files, and bin//obj/ from Docker build contexts.

GitHub Actions workflows (3 files)

• .github/workflows/docker-publish.yml — monthly GHCR publish of the CUDA / ROCm / OpenCL compiler-service images.
• .github/workflows/cleanup-ghcr.yml — monthly GHCR retention: keeps recent tags, prunes old ones so the container registry doesn't grow unbounded.
• .github/workflows/ci.yml — the canonical CI workflow is redesigned around the new compiler-service images. Jobs now cover: the standalone build, the xUnit test suite (run against the CPU backend), and GPU-bound Docker-run jobs that exercise the full pipeline inside a containerized CUDA/ROCm/OpenCL toolchain. This replaces the previous ILGPU.Tests.CPU/Cuda test harnesses.

Src/CLAUDE.md

An AI-assistant onboarding file at the top of Src/ describing:

• Build & test commands (including the --blame-hang-timeout 360s guard that prevents endless loops in compiler tests).
• Project structure at a glance (ILGPU / ILGPUC / ILGPUC.Compilers / ILGPUC.CompilerService / ILGPUC.Tests).
• The compilation pipeline from IL bytecode through ILFrontend → ModuleBuilder.Seal() → Optimizer → Backend.GenerateCode().
• IR architecture pointers (type hierarchy, value hierarchy, transformation framework).
• Pipeline-stage debug flags and snapshot-regeneration workflow.

Scope boundary

No test project (ILGPUC.Tests/), no samples (Samples/), and no backend / compiler source changes.

Depends on

PR #1578