Expose shears/div (Lamé) penalties through the whole stack

[balbasty]

Part of #25.

Thread the linear-elastic shears (Lamé mu) and div (Lamé lambda) penalty weights through flow_matvec/flow_diag at every layer:

• cpu-lib, lib (CPU tested), cuda-lib + new matvec_all/diag_all CUHOST launchers in cuda-impl (nvcc compile).
• bind-py binding + numpy/torch/cupy wrappers (torch autograd too) + fastfields.any.

Verified: operator symmetry under DCT2/DFT (numpy) and autograd gradcheck (torch).