Make rand_core optional

[pinkforest]

As proposed in #442 moves rand_core as optional feature

This still gets used via dev-dependencies and depending on feature selection tests uses API impl by proxy or mock.

[burdges]

I suppose not being default is justified by saying most dependencies already specify their own usages of the lower level methods.

[rozbb]

This is nearly ready. An annoying issue is that rand_core is actually enabled by default. This is because std enables rand_core/std. @tarcieri I don't think there's any way around this other than:

1. Kill rand_core/std and don't let anyone use it. The thing we lose here is only impl<R: RngCore> RngCore for Box<R>.
2. Make rand_core a default dependency, breaking our opt-in semver-breakage policy

I think (1) is the right choice. I don't think anyone will miss it.

[rozbb]

@tarcieri thoughts on killing rand_core/std? If we do this, we get the extra benefit of killing std entirely actually.

[rozbb]

Actually this can be solved with weak dependency features if we bump MSRV to 1.60. I think I'd be fine with that actually

[tarcieri]

AFAIK nothing in the crate itself actually needs or uses rand_core/std.

It provides two things:

• transitive activation of rand_core/getrandom feature which provides OsRng
• std::error::Error impl on rand_core::Error, which is not found in curve25519-dalek's public API

OsRng is used for testing in various places. It seems we can activate rand_core/getrandom in dev-dependencies and call it a day.

[tarcieri]

#458 removes activation of rand_core/std