Refactored sparse graphs to allow different backend implementations

[d-michail]

First pass over our sparse graphs support. Changes:

• Refactored sparse graphs to allow changing the actual implementation. This is needed since there are different tradeoffs.
• Simplified the undirected implementation which keeps source, and target matrices instead of keeping the transpose of the incidence matrix (which is really the same thing)
• Improved the directed implementation by allowing different handling of incoming edges (full support, lazy initialization or no support). This helps saving space if some algorithm does not need incoming edges.
• Added sparse graphs construction from streams which reduces copying.

---

• I read and understood https://github.com/jgrapht/jgrapht/wiki/Become-a-Contributor
• I read and understood https://github.com/jgrapht/jgrapht/wiki/How-to-make-your-first-%28code%29-contribution
• I added unit tests
• I added documentation
• I followed the Coding and Style Conventions
• I have not modified HISTORY.md or CONTRIBUTORS.md
• I ensured that the git commit message is a good one

[jsichi]

Could you merge from master? We had a problem with Travis CI, so pull requests weren't getting CI build runs.

[jsichi]

Since this is a brand new interface, would it make sense to use long instead of int with an eye towards GraphIterables?

[jsichi]

Just realized that my last comment was a little silly given the Integer type for vertex/edge. But I guess its better for uniformity in case we introduce a BigSparseSpecifics, and you've already made the changes, so merging now. Thanks!