Support generic data types in OpMeta

[jiazhihao]

FlexFlow currently assumes that the input/output tensors of some operators are 32-bit floating points. This issues tracks the progress of supporting general data types for FlexFlow's operators. The list of todos:

• Add input/output/weight data type as properties in OpMeta. This allows forward/backward tasks to know the data types of the input/weight/output tensors
• Use GenericTensorAccessor in forward/backward tasks

[lockshaw]

@jiazhihao The instructions above are out of date, especially after #622. Can you convert this into a list of which operators to which we still want to add multi-precision support?

[lockshaw]

Actually, let's move the list over to #408