v0.3.259-alpha

A large model split between your graphics card and your processor loads again on cards with
12 GB or more. On v0.3.258, Llama 3.3 70B on a 64 GB computer with a 12 GB card was refused:
the split put more of the model on the card than the loader lets it hold, so the loader left no
room for the conversation, moved the whole model to the processor, and the processor alone did
not have the memory. The split now leaves the card the share the loader keeps free, a layer or
three fewer on the card, and the model runs as it did on v0.3.256 (#364). The gap grew with the
card, and appeared in v0.3.257, when a split started taking the memory of idle models.

Reloading the config file now applies every setting in it. If you edit config.toml by
hand and reload it (Settings' reload, POST /api/admin/config/reload, or kill -HUP), only five
request and timing limits took effect; a changed contribution level, disk or graphics-card limit,
bandwidth cap or any other setting waited for the next time you saved Settings, though the reload
said it was applied. The reload now makes the whole file the running configuration, as a save does,
including how many processor cores the next model loaded uses and whether automatic model
management runs.

The setup wizard and the dashboard's disk meter show the drive your models are stored on.
They showed every drive's space added together, so on a computer with several drives the wizard
could suggest keeping more models than the drive holding them has room for, and the meter did not
match what your computer actually lets itself use. They now read the same drive the storage limit,
the download check and what other computers are told already use (#334).

---

Full commit list: v0.3.258-alpha...v0.3.259-alpha