Per-model reasoning memory is still missing across clients

[NikitaMGrimm]

#2324 is closed as completed, but current main does not implement the requested behavior on web/desktop or mobile.

Web / Desktop

Step	What happens	Result
1	Sol is selected with medium reasoning	Sol / medium
2	Switch to Luna	Luna / medium
3	Manually set Luna reasoning to max	Luna / max
4	Switch back to Sol	Sol / max

Web/desktop keeps one sticky ModelSelection per provider, so options carry across models instead of being remembered per model.

Mobile

Step	What happens	Result
1	Sol is selected with medium reasoning	Sol / medium
2	Switch to Luna	Luna / default
3	Manually set Luna reasoning to max	Luna / max
4	Switch back to Sol	Sol / default

Mobile rebuilds non-current models from their defaults.

Expected behavior

Step	What happens	Result
1	Sol is selected with medium reasoning	Sol / medium
2	Switch to Luna	Luna / default
3	Manually set Luna reasoning to max	Luna / max
4	Switch back to Sol	Sol / medium
5	Switch back to Luna	Luna / max

Mobile stores only the current ModelSelection, so switching loses the previous model's options.

I can also implement this across web/desktop and mobile with regression tests if needed.

Also a question that should be considered:
Should per-model memory apply only to reasoning effort, or to the full ProviderOptionSelection[].

I feel like it should not do the second thing, since "fast" should probably be a global toggle. Though, I can see the benefit of having it coupled to the model. (Having Luna always on fast for example)

[NikitaMGrimm]

One related behavioral question w/ #6508:

My expectation is that existing threads keep their own model/reasoning state unchanged.

Example (last used settings):

• Thread A: Sol / medium
• Thread B: Luna / max
• Reopening A should still be Sol / medium
• Reopening B should still be Luna / max
• Opening Thread C in the same project should inherit the last-used pair: Luna / max

So #6508 affects initialization of new threads, while per-model memory handles model-switching within a thread.