ci: raise the Live / Kind per-test budget to 150 s - #13960

Failure

Live / Kind on linux_amd64 fails its 120 s per-test budget intermittently. The three-agent OpenShift test took 112.3 s before #13947, then 123.2 s and 121.7 s in two of four runs after it added a runtime check to each agent refresh. About 30 s of every Kubernetes test's destroy waits out the sandbox pods' termination grace period: their processes catch SIGTERM but don't exit (#13959, NVIDIA/OpenShell#5025).

Decision

Raise the Live / Kind per-test budget and the live-kind profile's slow-timeout together to 150 s, as .config/test-budgets.yaml requires, and record the measurements and when to lower it. The step's wall budget stays at 270 s; the slowest failing run's nextest step took 217 s.

The test can't get faster in NemoClaw: the destroy wait is OpenShell's (NVIDIA/OpenShell#5025, the Kubernetes form of the not-planned NVIDIA/OpenShell#4635), and #13959 tracks lowering the budget once a fixed OpenShell is pinned.

Validation

• cargo nextest run -p nemoclaw-build: 204 passed, including the check that each step's slow-timeout matches its budget.
• Live / Kind on linux_arm64 (DGX Spark), measured while investigating: the three-agent test took 81 s, of which the first apply took 30–36 s and destroy 34 s; every sandbox's agent and supervisor pods stayed running until the 30 s grace deadline.
• Deferred to CI: Live / Kind on linux_amd64, where the budget was exceeded.