Allow the encoding to be generated on the fly in solve()

[garrison]

Summary

Fixes #9.

Details and comments

Remaining tasks:

• Add more test coverage
• Figure out a workaround for our prior need to access self.encoding.offset in solve(). offset somehow needs to make it from solve_relaxed() to solve(). It seems like solve_relaxed() almost needs to return the offset as a third tuple element ... or, we could put it in the RoundingContext, but it doesn't seem to really belong there. Potentially related to Provide access to rounding post-processing as a method #8. Possible solution in Deprecate encoding.offset while incorporating it into encoding.qubit_op #14.