Allow read_slice to return a Vec

[ia0]

Similar to #486 but takes the opposite approach:

• Memory-mapped storage can continue to return the slice directly
• Storage that require a mutable slice must allocate themselves

This is preferable to #486 because:

• Nordic has similar performance
• File-backed storage is not for embedded so allocating is ok

If we ever need to support embedded-storage trait, then we could add an Option<&mut [u8]> argument to be used and returned in the Cow<[u8]>. This would avoid allocation and copy. But since we don't have this need, we don't do it.

[coveralls]

Coverage decreased (-0.04%) to 89.907% when pulling 0f368f0 on ia0:read_cow into 85fe9cd on google:develop.

[kaczmarczyck]

I think I like this a bit better than #486. Might be a personal preference, but a quick comment on the non-intuitive use of Cow could be nice (we don't actually care about the writing part).

I find the binary size increases of both PRs surprising (in relation, at least). Would be great if we could check somehow if it's variance, but I don't see how.

[ia0]

Given that in #488 the binary size didn't increase while adding code, I'm assuming this PR is rather at the bottom of the current slot of added size.