also make a cleaner distinction between keychain, seed, path
This enables using `unsafe_prompts`, because with the original code, if
there was no namespace match, we wouldn't know which curve to use.
For ease of implementation, we use a LRU cache for derived keys,
instead of the original design "one cache entry per namespace".
SLIP21 is now treated completely separately, via `slip21_namespaces` and
`derive_slip21` method.
If more slip21-like things come in the future, we can instead hang them
on the keychain: put a per-curve Keychain object accessible by
`keychain[curve_name].derive()`, and the majority usecase will just pass
around `keychain[curve_name]` instead of having to specify the curve in
every `derive()` call.
Or alternately we'll just specify the curve in every `derive()` call,
whichever seems more appropriate.
This is to avoid including app-specific functionality in storage and
avoid circular imports. The following policy is now in effect: modules
from `storage` namespace must not import from `apps` namespace.
In most files, the change only involves changing import paths.
A minor refactor was needed in case of webauthn: basic get/set/delete
functionality was left in storage.webauthn, and more advanced logic on
top of it was moved to apps.webauthn.resident_credentials.
A significant refactor was needed for sd_salt, where application (and
UI) logic was tightly coupled with the IO code. This is now separated,
and storage.sd_salt deals exclusively with the IO side, while the app/UI
logic is implemented on top of it in apps.common.sd_salt and
apps.management.sd_protect.
