webapp: fix GPT storage

[tharvik]

as outlined in #800, the storage space of the browser does break when huge models (such as GPT) are stored.

• change the persist backend to support larger stores as the old one doesn't support theses
• add a storage backend based on OPFS
• a smaller storage backend based on LocalStorage, to still work if OPFS isn't supported (such as in the tests)
  • stores the smaller models until capacity, where it discards the large ones
• also add a layer to check that deserialized values are valid
• also reduce the serialized size of weights & models by using UInt8Array instead of array of numbers
• change the msgpack library to a one not eating up all memory (and recent, and standard)
• update server tests to wait for closing the server, and reduce the number of loaded task per test

[tharvik]

finally managed to find a way to avoid clogging up memory: don't trust msgpack-lite to do the job correctly and instead use the official JS library

[JulienVig]

Thanks for the great work!

Using Chrome or Firefox (131) smaller models are saved with LocalStorage but GPT never seems to be persisted (refreshing clears it), is there anything to do to enable OPFS?
Edit: that was my mistake

[JulienVig]

Great work!! I left mostly questions and pleas for more comments in the code, nothing blocking :)