Rmills/fix vecreplacearray seqcuda extra copy/release
This MR removes an unneeded copy from GPU in VecReplaceArray_SeqCUDA()
(it's not needed because the user has ceded control of the array to PETSc), and also adds a comment about why VecBindToCPU_SeqCUDA()
always sets the SeqCUDA
version for ops->replacearray
, rather than setting the Seq
version, as is done for every other operation.
This replaces MR !3400 (closed), which I will close.