Speed up ProDy

The patch replaces two functions (prody.dynamics.anm.ANMBase.buildHessian and prody.dynamics.anm.solveEig); calcANM/calcModes route through them. buildHessian is general and mathematically exact: it builds the Hessian in float64 unconditionally, supports AtomGroup/Atomic inputs (via _getCoords/getCoords) and raw numpy coord arrays (via checkCoords), any cutoff>0 (validated by checkENMParameters), and both scalar gamma and callable Gamma objects (callable invoked as gamma(dist2,i,j), matching upstream exactly, incl. squared-distance arg). It also restores a real sparse CSR Kirchhoff matrix and attaches the build coords to the sparse Hessian instance (no global state). solveEig has a deterministic eigsh shift-invert path (eigsh sigma=-1e-8, which=LM, tol=1e-10) that handles any sparse M with a finite integer n_modes < dof for both zeros=False and zeros=True, with an internal guard that defers to upstream if final_n_modes exceeds available values. The LOBPCG accelerated path (the fast path actually measured) is taken only for the standard ANM config: M sparse, reverse=False, integer n_modes < dof, zeros=False, expct_n_zeros==6, build coords attached, coords.shape[0]*3==dof, and coords.shape[0] < 12000; its result is accepted only after a residual-norm gate (max scaled residual <= 1e-3) and otherwise falls through to the deterministic eigsh path. Benchmark tiers 1OEL/5GAR/1aon/6TLJ (3668–9200 Cα) exercise LOBPCG; the 7B0U OOD tier (13320 Cα) is >=12000 so it uses the eigsh shift-invert path, not LOBPCG.