smu10_smu_init error paths leave hwmgr->smu_backend dangling after kfree(priv)

Summary

smu_backend is published before the two BO allocations; both error paths kfree(priv) without NULLing it (and mask the real error with -EINVAL). NOT reachable today (status.sw gates sw_fini dispatch) - latent lifetime trap for any future error-path unwind. Requires Raven silicon. NULL-out + propagate-real-error one-liner fix.