Avoid two clones in 'reduce'.

[lemmih]

Hi,

reduce needlessly clones the numerator and denominator. The "proper" way to fix this would be to add a new trait bound but that is not backwards compatible. This PR proposes a temporary fix: Take the values and leave behind zeroes. This avoids cloning and T::zero() usually doesn't involve heap allocations (at least they don't for BigInt). All in all, this PR improves performance by about 10% (1150ns -> 1050ns) when creating values from small BigInts (tested on an M1).
Creating a Ratio<BigInt> is still about 10x slower than Ratio<isize> but that is entirely due to problems in num-bigint.

Note: I copied the rng module from num-bigint.

Note: Adding a NumAssignRef bound would get rid of g.clone().

[cmpute]

There are also many other functions that use clone(). I will also vote for adding NumAssignRef or NumRef bound to the type. Although it will be a break change, it's quite worth it

[cuviper]

@cmpute

Although it will be a break change, it's quite worth it

I don't take breaking changes lightly. Maybe we could add those bounds with a semver bump, but not right now.

[cuviper]

Thanks for the PR! I moved the benchmarks to a CI crate so rand doesn't look like a public feature, nor affect MSRV as a dev-dependency (which can't be optional). I also added the same replace-zero trick to the final sign normalization.

bors r+

[bors]

Build succeeded:

• Format
• No Std (1.31.0)
• No Std (stable)
• Test (1.31.0)
• Test (1.36.0)
• Test (beta)
• Test (nightly)
• Test (stable)