If not runtime-writeable weights, then we can still minimize the accumulator bit width according to the data types.

InsertDWC now uses hls mode, and vivado mode is only used for
8-bit divisible stream widths

- do not set in/outFIFODepths attrs to 0 after setting
- do not set in/outFIFODepths attrs to max() in InsertFIFO
- special handling of first/last FIFO depths for characterization