refactor: hybrid incremental solver and performance improvements

[baszalmstra]

This is a big PR that does two things:

Incremental solver

The solver is now incremental by default. This means that the solver will only request information from the DependencyProvider when it is relatively sure that that information is needed to come up with a solution. Previously the solver would greedily request all information about the problem space up front. This made the solver very unsuitable for cases where requesting information about packages is expensive.

Optionally a DependencyProvider has the ability to hint to the solver that it already has dependency information available for some candidates. The solver will use this information to increase its knowledge of the problem space without requiring network operations. The latter is used for the conda solver.

In general, it's better to have all the information available upfront because if the solver has "all" the information available it is able to reach a solution quicker because it doesn't have to guess about information it doesn't know.

Performance improvements

This PR also implements a performance improvement by being more thoughtful about making decisions about the next steps to take in the algorithm. Instead of picking the first available decision it can make it checks all possible choices to make and picks the decision that involves the least amount of versions to pick.

This is similar to what pubgrub does:

https://github.com/dart-lang/pub/blob/master/doc/solver.md#decision-making

In the future, we might take a look at different heuristics but this implementation is good enough for now.

Some results:

	libsolv	libsolv-rs 0.7.0	libsolv-rs (this)
python=3.9	7.03ms	3.47ms	3.57ms
xtensor, xsimd	5.28ms	2.00ms	2.22ms
tensorflow	773.27ms	407.98ms	152.94ms
quetz	1380.8ms	1684.7ms	301.53ms
tensorboard=2.1.1, grpc-cpp=1.39.1	515.25ms	122.48ms	89.41ms
rubin-env	7459.7ms	3633.2ms	1798.8ms

Note that for "simple" cases this PR is slower but just slightly and for "complex" cases this PR is much faster.