Fix GPU CI: disable codecov safe.directory write (.gitconfig lock race)

[ChrisRackauckas-Claude]

Problem

The GPU Tests workflow on main is intermittently red. Its two matrix jobs — CUDA Core and CUDA Applications — run concurrently on the same self-hosted GPU runner host, sharing a single /home/chrisrackauckas/.gitconfig.

The codecov/codecov-action@v7 post-test step runs git config --global --add safe.directory .... When both matrix jobs reach that step at overlapping times, they race on git's config-file lock and one job dies with:

error: could not lock config file /home/chrisrackauckas/.gitconfig: File exists
##[error]Process completed with exit code 255.

This happens after ComplementaritySolve tests passed is printed, so the tests themselves are fine — it is purely the codecov action's git-config setup racing. Which of the two jobs loses the race varies run-to-run (it hit CUDA Applications on the latest main run and CUDA Core on a recent PR run), confirming a concurrency race rather than a code/test issue.

fail_ci_if_error: false is already set, but it does not help here: the failure is in the action's safe.directory setup step, which exits 255 before that flag is consulted.

Fix

Set disable_safe_directory: true on the codecov step. This skips the git config --global --add safe.directory write, which is the only operation racing on the shared ~/.gitconfig. The write is unnecessary on these self-hosted runners (the workspace is already owned by the runner user), so disabling it removes the race without masking any failure. The test step (julia-runtest) still fully gates the job, and coverage still uploads.

Notes

• Workflow-only change; no .jl files touched (no Runic formatting needed).
• This is a separate concern from the Downgrade (QA) precompile-segfault fix (Fix Downgrade (QA): remove PATH solve from precompile workload (intermittent libpath50 segfault) #66); kept as its own focused PR.
• I confirmed from the CI logs that the failing step is codecov-action's safe.directory setup and that the tests pass before it; this is a self-hosted-runner-environment race, so it cannot be reproduced on a non-GPU dev box.

Please ignore until reviewed by @ChrisRackauckas