Guard against slow GPUCompiler versions

[ChrisRackauckas-Claude]

Fixes #434.

Summary

• Add GPUCompiler as a weak dependency with a compat floor of 1.13.2.
• Add a QA guard that checks the weakdep UUID and compat floor stay in Project.toml.
• Add the small QA test dependencies needed for the guard.

Root cause

The adaptive Lorenz regression tracks to a GPUCompiler/NVPTX codegen cliff in GPUCompiler 1.13.1 and earlier. On the stock DiffEqGPU adaptive kernel body, GPUCompiler 1.13.1 produced the slow path at about 211.8 ms; GPUCompiler 1.13.2 restored the expected path at about 47.9 ms.

The minimal compiler-pipeline difference I isolated is the EarlyCSEPass() inserted after SROAPass() in GPUCompiler's early simplification pipeline. Rebuilding GPUCompiler 1.13.1 with only that pass restored the adaptive benchmark to about 47.8 ms, matching 1.13.2+ behavior.

This package-level fix avoids known-bad compiler versions without adding GPUCompiler as a hard dependency for non-GPU installs.

Validation

• JULIA_DEPOT_PATH=$PWD/dev/julia_depot: GROUP=QA julia --project=test test/runtests.jl
  • Quality Assurance | 20 pass / 20 total
  • JET static analysis | 75 pass / 75 total
• JULIA_DEPOT_PATH=$PWD/dev/julia_depot: julia --project=test -e 'using Pkg; Pkg.status(["CUDA", "GPUCompiler"]; mode = Pkg.PKGMODE_MANIFEST)'
  • CUDA v6.2.0
  • GPUCompiler v1.22.7
• JuliaFormatter.format run on the changed files.

Notes

The GPU benchmark investigation was performed on demeter2 with the issue reproducer. Other candidates checked and ruled out included extra rejected steps, saving/output, tstops, DAE init/finalization, workgroup size, Julia version, fixed-step Tsit5, raw SVector Lorenz loops, and the stock DiffEqGPU step! body itself.