performance optimization

[liminchen]

• Eigen dynamic-size matrix -> static-size;
• Add std::vector isFixedVert to avoid binary search on std::set fixedVert;
• Local update for mesh features and Hessian matrix data structure rather than recompute whole;
• Compute Hessian and gradient together so that certain variables can be shared;
• Improve multi-threading
  • Parallel energy computation
  • Parallel gradient entry computation and vector assembly
  • Parallel Hessian matrix assembly
  • other code that's reasonable to be parallelized...
• SIMD optimization
• GPU optimization
• Algorithmic improvement (see our paper for more detail)