Simplify the computeOperandLatency API.

The logic for recomputing latency based on a ScheduleDAG edge was
shady. This bypasses the problem by requiring the client to provide
operand indices. This ensures consistent use of the machine model's
API.

llvm-svn: 162420