balance: Remove MakeBalance::from_rng

[olix0r]

MakeBalance's use of SmallRng is problematic: since it clones the
SmallRng between Balance instances, each instance will have the same
state and produce an identical sequence of values. This probably isn't
dangerous, though it is certainly unexpected.

This change removes the MakeBalance::from_rng and
MakeBalanceLayer::from_rng helpers. The MakeBalance service now uses
the default RNG via Balance::new. Balance::new now creates its
SmallRng from the thread_rng instead of the default entropy source,
as the default entropy source may use the slower getrandom. From the
rand docs:

In case the overhead of using getrandom to seed many PRNGs is an
issue, one may prefer to seed from a local PRNG, e.g.
from_rng(thread_rng()).unwrap().

Finally, this change updates the balancer to the most recent version of
rand, v0.8.0.

This seems like a change we should make before the next release/version
bump of tower.

[hawkw]

Looks like the examples need to be updated.

[hawkw]

This seems good to me once the examples compile again!

It does seem like it could be possible to allow constructing MakeBalance instances with an overridden RNG, perhaps by adding a constructor that takes an FnMut() -> Rng? That way, the RNG is overridable at the MakeService level as well as the Service level, but with a way of making a new RNG with its own state. But, we can add that in a follow-up if we decide it's even worth having. In practice, I'm not really sure if there's all that much of a need for this --- maybe if you want to mock out randomness for deterministic testing? I dunno.

[olix0r]

@hawkw We still support Balance::from_rng, so tests that use an actual balancer are still feasible. I don't think it's really worth supporting this on the Make/Layer at this point (especially if we have to support an additional generic).

[hawkw]

@hawkw We still support Balance::from_rng, so tests that use an actual balancer are still feasible. I don't think it's really worth supporting this on the Make/Layer at this point (especially if we have to support an additional generic).

that's fair, this is fine!