Indeed, by using the Env.nset and Env.nget defined in
0d6575a7 a pb that was hided by the
use of Env.set and Env.get was triggered.

The bug was that variables initialisations were done several times.

In sasa.ml, the way I was updating the Env.t by iterating on the
variables was forcing to "close" expressions that were calling init
functions.

More precisely, they were not consistent by the behavior of -custd mode.
Indeed, It was returning the values of variables after the step is performed.

In the custom demon mode, sasa returns Enab and variable values before
the Activate is computed by the demon.

Chosing one behavior or the other is arguable ; the main advantage of
this choice is that we don't miss the initial values.

The rationale is that
 - it is not always the cmxs that is necessary .e.g., rdbg uses .cma or .ml
 - it makes more sense to refer to the source code

This can be used from rdbg from example, to generate session files.

In particular, 'rdbg -lurette' now works thanls to this option
(cf rdbg sha 63566a893807361691137cafe15854a2a7395eec)

Remove the --generate-rdbg-load one that is less general

Save/restore the PRGS at checkpoints (necessary for internal demons).

This is necessary because I cannot Dynlink .cmxs files from a ocaml
toplevel. Hence, one way to load the algos from rdbg is to
#load .cma files. This option generates an ocaml file that
can be #used by automatically generated rdbg-session.ml files.

Rationale: it was not really necessary (Algo.neighbor contains almost
the same information), and the Topology.neighbor type was alone in its
module, which looked weird.