Move constrained logic to a special ConstrainedHilbertIndex class that can more easily be overridden (#1528)

> I believe there were a few requests for improvements of the hilbert
spaces, but none of them would really benefit from jax afaik. In
particular:
>
> - make restricted fock indexable in more cases (e.g. fixed particle
number is small but we have a lot of sites). Implemented it
[here](https://github.com/inailuig/chemistry_nqs_second/blob/5c04364aa456071a2a07a9f57ef7fd91709aa377/constrained_fock.py)
some time ago and can try to clean it up for a PR at some point. Only
really useful for ED and exactState (will benefit [@Jannes
Nys](https://cqslab.slack.com/team/U02FE43BC9G) and [@Fan
Yang](https://cqslab.slack.com/team/U040MJYMZ3L))
>
> - restricted fock random_state going OOM if the local cuttoff is big
and the local fock space is therefore not really indexable (according to
[@Zakari Denis](https://cqslab.slack.com/team/U043EP0T1DM))
>
> - I would change the default dtype of the discrete hilbert spaces to
something more reasonable like int8/uint8 instead of double. I think
it's cleaner to set it in hilbert already instead of later casting in
the samplers. Also types are currently not consistent so you might get
sth different from all_states and random_state

This is a draft to make it easy to implement point 2, essentially we
could add some _special_ HilbertIndexConstrained for what z-denis
wanted, I think.
You could take this as a starting point.

We can probably further clean the interface, but I did not think much
about it.

Then inailuig step 1 could be achieved by jaxifying those classes.