Fixing some implicit int64->32 downcast errors. (In most cases, by

having the ops explicitly return a failure that they can't handle
overly-large inputs).  Most of these should never affect correct
tf programs until people get a lot more memory in their machines.
Change: 118476613