Support ?Sized types in Rng*: TryRng* blanket impls

[tarcieri]

Adds ?Sized to the bounds of the blanket impls, so unsized types can also be used by way of the blanket impl.

See also: #45

[baloo]

Downsteam changes with that:
RustCrypto/SSH#453

Without:
RustCrypto/formats#2175

[tarcieri]

Well this is interesting:

https://play.rust-lang.org/?version=stable&mode=debug&edition=2024&gist=218bb58ea0812eb2f347d8b5008166ab

It seems like if it were:

pub trait CryptoRng: RngCore {}

instead of

pub trait CryptoRng: TryRngCore<Error = Infallible> {

...then the entire problem just goes away without having to add an explicit ?Sized bound on the blanket impl?

Edit: aah nope, I think it still needs this change as well to be able to work with trait objects, and we still need a TryCryptoRng bound for CryptoRng, but with this change it's possible to add an explicit RngCore bound to CryptoRng, like:

pub trait CryptoRng: RngCore + TryCryptoRng<Error = Infallible> {}

...which will ensure/test that a CryptoRng can always act as an RngCore (and oddly without this change it breaks the tests and I'm not sure why).

I will save that for a followup though, so as not to muddy the issue, unless everyone likes it.

[tarcieri]

@baloo yeah, we use <R: CryptoRng + ?Sized> in several places, so it's going to be really annoying (and ideally not necessary) to change all of those to add a RngCore bound

@dhardy I don't suppose it would be possible to get another prerelease out with this PR (and possibly with the added CryptoRng: RngCore bound)? It's going to be annoying to update to v0.10.0-rc-4 as-is

[baloo]

Yeah it's a much better solution than having to add the RngCore to every consumer. An RC would really be appreciated!

[tarcieri]

I went ahead and did the bounds change I proposed above in 9966877:

pub trait CryptoRng: RngCore + TryCryptoRng<Error = Infallible> {}

It's just more explicit and ensures that if a CryptoRng works at all it will work as an RngCore.

I kept it in a separate commit in case you just want the ?Sized change for a stopgap rc

[dhardy]

I have no problem with this change. I'll make a new release later (not much time now).

[dhardy]

Interesting that this is necessary; it seems to indicate a trait solver issue.

But simple solutions are good!

[tarcieri]

I think having explicit bounds even if they're covered by a blanket impl makes the compiler errors much better, because it will tell you explicitly why the bound wasn't satisfied, whereas figuring out why a blanket impl didn't work can be a lot of red herrings