Merge RngCore, TryRngCore

[dhardy]

• Revise traits
• Update documentation
• Added a CHANGELOG.md entry

Summary & motivation

It was noted in #13 that the impl relationship between RngCore and TryRngCore is unsatisfactory. @newpavlov proposed a revision where RngCore: TryRngCore. This PR goes a bit further, merging the two traits.

• Delete RngCore
• Rename TryRngCore → TryRng. (Opinionated and not required. I also considered renaming it to RngCore.)
• Add trait InfallibleRng: TryRng<Error = Infallible>. This is effectively a trait alias with no methods of its own.
• Merge CryptoRng and TryCryptoRng into CryptoRng: TryRng

Results (motivation):

• One/two less traits (depending on whether you count InfallibleRng as a "trait" or "trait alias")
• Solves the issues surrounding impls on RNGs; in particular &mut R: TryRng for R: TryRng
• UnwrapMut is no longer needed

(And yes, I can see the hypocrisy of suggesting these breaking changes after complaining about others, but preparing for rand_core v1.0 means deciding whether to deal with existing issues or stabilise what we already have.)

Optimality

In theory there should be little impact on the generated code, but it's clear the result isn't identical. I ran all rand benchmarks against master; those with at least 5% time change:

random_u32/chacha20: -8%
random_u32/std: +5%
random_u32/thread: +7%
random_u64/thread: +7%
init_gen/pcg64: +5%
init_gen/std: +11%
init_from_u64/chacha12: -5%
init_from_u64/chacha20: -5%
init_from_u64/std: -7%
init_from_seed/pcg64mcg: -9%
seq_slice_choose_1_of_100: +101%
seq_slice_sample_weighted_1_of_1000: +8%
seq_iter_sample_10_of_100: +24%
seq_iter_sample_fill_10_of_100: +33%
choose_unhinted_from_1_ChaCha20: -5%
choose_windowed_from_1_ChaCha20: -5%
choose_size-hinted_from_2_ChaCha20: +7%
choose_size-hinted_from_3_ChaCha20: +7%
choose_size-hinted_from_10_ChaCha20: +6%
choose_size-hinted_from_100_ChaCha20: +7%
choose_size-hinted_from_1000_ChaCha20: +7%
choose_windowed_from_1000_ChaCha20: +8%
choose_windowed_from_1_Pcg32: +9%
choose_windowed_from_100_Pcg32: +8%
choose_windowed_from_1000_Pcg32: -8%
choose_size-hinted_from_2_Pcg64: -9%
choose_size-hinted_from_3_Pcg64: -9%
choose_size-hinted_from_10_Pcg64: -9%
choose_size-hinted_from_100_Pcg64: -9%
choose_size-hinted_from_1000_Pcg64: -9%
choose_windowed_from_1000_Pcg64: +7%
shuffle_2_ChaCha12: -6%
partial_shuffle_10000_ChaCha12: +5%
shuffle_2_Pcg32: +83%
sample_i8x1/Pcg64/single: -9%
sample_i16x1/SmallRng/single: +11%
sample_i16x1/Pcg32/single: +6%
sample_i16x1/Pcg64/single: +7%
sample_i32x1/Pcg32/single: -12%
sample_i32x1/Pcg64/single: -12%
sample_i64x1/ChaCha8Rng/distr: -7%
sample_i64x1/Pcg32/distr: -8%
uniform_single/f64/Pcg32: +5%
uniform_distribution/f32/SmallRng: +10%
uniform_distribution/f32/Pcg32: -7%
uniform_distribution/f64/Pcg32: +19%
weighted_index_modification: -15%
weighted_sample_indices_10_of_1k: -6%
weighted_sample_indices_100_of_1k: -10%
weighted_sample_indices_400_of_1M: -6%
weighted_sample_indices_600_of_1M: -6%
weighted_sample_indices_1000_of_1M: -10%

The majority of results have no significant change and the above have roughly as many wins as losses, though there are a few more significant losses. TODO: look into this more.

Alternatives

1. No change. We commit to keeping UnwrapMut and cannot have &mut R: TryRngCore for R: TryRngCore.
2. We revert rand_core: add blanket impl of TryRngCore for RngCore rand#1499. Infallible RNGs need to impl both RngCore and TryRngCore (possible via macros).
3. @newpavlov's suggestion. Infallible RNGs need to impl TryRngCore<Error = Infallible> (same as this PR). dyn RngCore vtables have extra methods.
4. A variant on @newpavlov's suggestion / this PR where RngCore (or InfallibleRng) derives TryRngCore<Error = Infallible> + Sized. This implies a need to use dyn TryRngCore<Error = Infallible> instead of dyn InfallibleRng, but also that the vtable will not feature both try_next_u32 and next_u32.
5. With the (long time unstable) trait-alias feature we could actually make InfallibleRng a trait alias. (I don't know whether switching to this this once stable would be a breaking change.)

Details

See also:

• Draft: rand_core::TryRng getrandom#772
• Draft: rand_core::TryRng rand#1706

Concerns

Would it be preferable to stabilise this code in rand_core v1.0 over what we have now? I think so but would very much like feedback on this. CC @tarcieri @baloo

How is usability of the result?

• Implementing an InfallibleRng requires a tiny bit more boilerplate.
• Testing / using InfallibleRng without rand::Rng is a slight pain without the infallible methods next_u32() etc.; one must either .unwrap() (good for tests) or use match to unpack the Result. In PRNG crates this is a small (IMO acceptable) burden.
• Usability of rand::Rng is basically unaffected
• Is this a big deal for rand_core users, e.g. RustCrypto?

How much work is involved in dealing with the changes?

• It's a breaking change to core traits, temporarily requiring a compatibility-bridge like rand_core_compat for some applications.
• Updating RngCore implementations to use TryRng with type Error = Infallible and different method prototypes is a minor PITA. This affects us more than users; updating the rust-random/rngs repo will be a pain.
• Code using rand::Rng is basically unaffected, except that usages of RngCore must be replaced with Rng or InfallibleRng
• Code using rand_core::RngCore infallible methods (next_u32 etc.) without rand::Rng has to find some alternative