Tpetra::MultiVector::norm*: Prefer running on device
Created by: mhoemmen
Tpetra::MultiVector::norm* methods currently run where the MultiVector's data were most recently updated. Reports from @vbrunini running SPARC with vectors with 1.8 million local rows show that this can cause a slowdown, if running on host. We need to fix these methods so they prefer running on device unless the vector has too few rows to justify running on device.