Deslop wedged-VF detection: dedupe sentinel report guard, trim redundant tests

[yummybomb]

Cleanup pass on #435, stacked on hypeship/vgpu-wedge-quarantine. No behavior changes to the detection path; net -130 lines, mostly tests.

Production code

• vgpu_sentinel.go: handleFailure/handleSuccess duplicated the confirm-assignment preamble (confirm → warn on error → skip on change). Folded the logging into confirmAssignment, which now returns a bool and tags its log lines with an action attribute. Logging-only difference: the success path now also logs at info when it skips a changed assignment (previously silent).
• gpu_watch.go: moved the guestServer.GetGPUInitStatus RPC method out from between two gpuInitReporter methods so the reporter methods are contiguous. Pure move.
• process_identity.go / storage.go: trimmed two comments (one explained the diff rather than the code, one enumerated its callers).

Tests

• vgpu_sentinel_test.go (-177/+... net ~-120):
  • The fake reportFailure re-implemented the health store's per-assignment dedup, and ReportsFailureOnce then asserted that fake's own behavior. Dedup is the store's contract and is covered in lib/devices; the fake is now a plain recorder and the test (renamed ReportsFailure) keeps the UNKNOWN-not-reported and report-fields coverage, absorbing the NVRM log assertions from the deleted LogsNVRMMessageOnQuarantine.
  • Dropped PollsTargetsConcurrently — it exercised errgroup.SetLimit, not sentinel logic.
  • RepairsHealthStoreOncePerPoll proved once-per-poll with 64 targets; 2 targets prove the same thing.
  • Merged the three changed/released-assignment tests into table-driven ConfirmsAssignmentBeforeReporting covering all four state×assignment combinations (the failed×changed case is new).
  • Removed a brittle log-content assertion from RetriesFailedTallyClear.
• vf_health_test.go: the no-op-success test spun 64 goroutines, but nothing is contended — the assertion (no persist retry on a failed store) holds with one call. Kept its unique repair assertion (exactly one persist, two dir syncs).
• gpu_watch_test.go: folded success-is-terminal into TestGPUInitReporterState and dropped ProbeGPUInitSkipsWithoutNvidiaSMI (asserted only an empty log buffer on a trivial early return).

Deliberately not changed

• Constructor nil-checks and the vgpuSentinelStore downcast in NewVGPUSentinelController — same pattern as NewHealthCheckController, so they match package idiom (kept TestNewVGPUSentinelControllerRejectsUnsupportedManager with them).
• create.go setting GPUProfile outside setStoredVGPUDevice — deliberate: clearStoredVGPUDevice keeps the profile for the next start.
• All fail-closed semantics (terminal success, assignment revalidation, no per-report persist retry, uncertain-liveness preservation).

Testing

go test -race ./lib/system/guest_agent ./lib/guest ./lib/devices -count=1
go test -race ./lib/instances -run 'Test(VGPUSentinel|ListVGPUSentinel|ProbeGPUInit|RunGPUProbeAttempt|GPUInitReporter)' -count=1
go test ./lib/devices ./lib/providers
go vet ./lib/system/guest_agent ./lib/guest ./lib/instances ./lib/devices ./lib/providers

All pass locally (embedded VMM/agent binaries stubbed for compilation, same constraint as the parent PR).

[yummybomb]

folding these changes directly into #435 instead