Port GDN kernels to A5/950

[learning-chip]

Except for solve_tril phase that needs a re-design, other phases all run correctly on-device. Still uses the slow GM workspace, thus performance not faster yet.

TODO:

• Use fast C-V datapath like Idealized C-V data exchange kernels for A5 pto-kernels#174