some speedup for rendering sparse FD derivatives

Some changes for rendering sparse FD derivatives have been committed (especially matters for badly-vectorized probs), speedup about 10% has been observed on probs with 200-400 vars x 200-400 cons, each con with about 4-5 dependency vars (ipopt was involved). The changes affect sparse linear, nonlinear probs and sparse SLE.

It is recommended to retest your problems where nVars >> 1 with parameter useSparse = True|False|'auto'.