fix: don't abort install.sh when the OpenShell gateway health-check fails

[vyncint]

The bug

install.sh runs check_openshell before it installs the driver and runs setup:

main() { … check_openshell; resolve_version; install_driver; maybe_setup; }

check_openshell shells out to NVIDIA's official OpenShell installer, which starts and health-checks the gateway. But that gateway can't become reachable until this driver's setup wires a compute driver into it — which happens two steps later. So the OpenShell installer exits non-zero (local gateway listener did not become reachable within 30s), and because install.sh runs under set -eu, that aborts the whole script before setup ever runs.

Result of a fresh curl … | install.sh | sh: apple/container and OpenShell get installed, but the driver is never set up. The gateway then crash-loops (no compute driver configured / failed to connect to /tmp/oshl-ac/driver.sock) — exactly the broken state a fresh install just produced on my machine.

The fix

Tolerate the OpenShell installer's expected non-zero exit, but assert the binary actually installed, then let the script continue to install_driver → maybe_setup:

-       curl -LsSf "$OPENSHELL_INSTALL_URL" | sh
+       curl -LsSf "$OPENSHELL_INSTALL_URL" | sh || true
+       need openshell || err "OpenShell installation failed"

setup brings the gateway up at the end (it's idempotent and also repairs the CLI's gateway registration). A genuine OpenShell install failure is still caught — need openshell fails and we err out.

Verification

• sh -n install.sh parses; shellcheck clean (same set the CI job lints).
• Root cause reproduced live on a real machine: a fresh curl … | sh left the gateway crash-looping with no compute driver configured; running openshell-driver-applecontainer setup by hand fixed it completely (gateway Connected, sandbox create → Ready → exec uname -a returns Linux … 6.18.15 aarch64). This PR removes the need for that manual step.