Replace pre-trained weights for the Primal-Dual example (ckp_PDNet.pth)

For some reason, current pre-trained weights give very bad reconstructions on Learned Primal-Dual algorithm for CT scan (https://deepinv.org/auto_examples/unfolded/demo_learned_primal_dual.html#sphx-glr-auto-examples-unfolded-demo-learned-primal-dual-py).
The weights I propose are trained on the Primal-Dual example notebook, with 9 epochs.

ckp_PDNet.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c91402f71e3d5de9fdc3b45d789db597da8a50ae9c42108ff3e7457cfdf06086
+oid sha256:0e309b1433e5ced5d93b086980684309d27bca0e17ce9db5a4af7b349e177f8c
 size 3222941