Well, the startup delay being discussed here is a result of a MC Client copying the Library from the MC Server. So a smaller Library should improve that, however you make the Library smaller.
If you use the Waveform, and store the information in the file tag but not in the Library, you would see other performance issues as MC would have to read from the file whenever that data was accessed. I guess if you are playing the file, that wouldn't be an issue. But if any View touched it, or the Tagging Window, then MC would have to wait while it read the tag from the file. I don't know if MC is even capable of only storing the data in the file tag, and not in the Library. I would think currently that it could not, so there could be development issues there.
The data appears to be plain text at the moment, so any compression would help, at the cost of processing power to compress and decompress it, which isn't much.
Having the data in each file, and hence having larger files, would only be an issue if the drive they were on was nearly full. Drives perform much worse when they are nearly full. But the data isn't that larger, so that is unlikely to be an issue.
The real answer for me would be to have the file analysed in real-time and displayed only if the feature was turned on. That would require more processing power in the Client, but maybe not much more. With that solution neither the Library nor the files would be larger. But issues such as seeking in a file would need to be addressed. The waveform may lag the audio a little, which probably wouldn't be appreciated. But that entirely depends on the implementation.