Fix OptimizationProblem for SVector/SArray: use out-of-place form

[ChrisRackauckas-Claude]

Summary

This PR fixes a precompilation error that causes all tests (including CUDA) to fail.

The Problem

The newer SciMLBase enforces that immutable types like SVector/SArray must use the out-of-place OptimizationProblem{false}(...) form since they cannot be mutated in-place.

The error occurred during precompilation:

ERROR: LoadError: Initial condition incompatible with functional form.
Detected an in-place function with an initial condition of type Number or SArray.
This is incompatible because Numbers cannot be mutated, i.e.
\`x = 2.0; y = 2.0; x .= y\` will error.

If using a immutable initial condition type, please use the out-of-place form.

The Fix

All OptimizationProblem calls using SVector/SArray initial conditions are updated to explicitly use the out-of-place form OptimizationProblem{false}(...).

Files changed:

• src/precompilation.jl - main precompilation workload
• test/gpu.jl - CUDA tests
• test/regression.jl - regression tests
• test/reinit.jl - reinit tests
• test/constraints.jl - constraint tests
• test/lbfgs.jl - LBFGS tests

Fixes: https://github.com/ChrisRackauckas/InternalJunk/issues/26

[ChrisRackauckas-Claude]

GPU OOM Analysis

The GPU tests are failing with Out of GPU memory on the self-hosted runners. After investigation:

Root cause: The shared GPU runners (arctic1-2, arctic1-3, arctic1-4) have oversubscribed T4 GPUs (14.5 GiB). Other workloads consume most GPU memory, leaving insufficient memory for CUDA kernel JIT compilation.

Key findings:

• The CUDA memory pool shows only ~24 KiB allocated by our tests — the OOM is not from test array allocations
• The OOM occurs during cuLaunchKernel (hybrid test) or even cuDevicePrimaryCtxRetain (when GPU is completely occupied)
• GPU tests were already failing on main before this PR — with a different error: "Initial condition incompatible with functional form" during precompilation (the exact bug this PR fixes)
• CI (CPU tests) and format check pass consistently

Changes in this push:

• Added GC.gc(true) between test file includes to reclaim memory between tests
• Added CUDA.reclaim() at start of lbfgs.jl for when GPU memory is reclaimable
• These help when the runner has moderate memory pressure, but cannot fix completely oversubscribed GPUs

Recommendation: The GPU OOM is an infrastructure issue, not a code issue. This PR should be mergeable based on CI (CPU) + format check passing. The GPU runner capacity may need to be addressed separately.

[ChrisRackauckas-Claude]

CI Update — 26/27 GPU Tests Pass

After switching back to T4 runners (gpu-t4), GPU tests are mostly passing:

Test Suite	Result
CI (CPU)	✅ All pass
Format check	✅ Pass
CUDA optimizers (21 tests)	✅ All pass
CUDA constraints (5 tests)	✅ All pass
CUDA hybrid optimizers (1 test)	❌ MISALIGNED_ADDRESS during kernel compilation

The hybrid test failure is a CUDA error: misaligned address (code 716) during CuModule loading — this happens during JIT compilation of the gpu_simplebfgs_run! kernel, not at runtime. This appears to be a CUDA driver issue on the T4 runners, not related to the SVector/SArray fix.

Note on V100 runners: V100 (compute 7.0) is incompatible with CUDA 13+ drivers on the self-hosted runners. JULIA_CUDA_VERSION is deprecated and LocalPreferences.toml pinning doesn't help since the check is driver-level. T4 (compute 7.5) is the only viable GPU option.

[ChrisRackauckas]

This is a Base issue. JuliaGPU/CUDA.jl#3034 JuliaLang/julia#61154 is the fix for the test which should be released soon enough.