use abstraction layer for linear system solvers, perform QR-based preprocessing

[chriscoey]

this is a major PR with changes across the board

• create abstraction layer for linear system solvers
• add naive dense linear system methods
• cleanup tests and test different linear system methods
• enforce requirement that primitive cone indices be unit ranges (just requires that rows of G and h be reordered)
• add indirect linear system solver (implementation not optimized)
• add preprocessing functions (optional; and try to reuse one qr factorization for QR+Chol) to pass MOI tests; originally in WIP: preprocessing using QR factorizations #73
• reuse QR decompositions inside preprocessing function and also optionally for creating a QR-based linsys cache
• enable MOI tests to be sparse
• refac 2 QR-based linsys caches
• change examples to just return data (not preprocessed) rather than an optimizer object, and modify tests accordingly
• restructure tests to be like MOI tests: one function per model
• add native interface tests for preprocessing
• debug strange failing MOI test for dense A (fixed by always stepping in interior direction when finding initial iterate)
• get hessians more efficiently
• create function for doing check, preprocess, solve all in one (like optimize function in MOI already does)

[chriscoey]

all MOI linear and conic tests are now passing! (bar SDP, logdet, rootdet for which we must wait for jump-dev/MathOptInterface.jl#531)

[chriscoey]

all tests passing! some example tests are currently commented out because they are too large/inconvenient to run often