jacobian: do not rescale output by sqrt samples

[PhilipVinc]

Fixes #1613 .

Ensures that jacobian does not return a rescaled output unless an internal flag is specified.
Eventually this flag will be removed, but keeping it for now because it makes our life easier and would complicate @inailuig 's life and PR.

I should add some tests testing jacobian specifically...

[gcarleo]

Looks like it's breaking dynamics

[inailuig]

I support this change.

We could multiply the intermediate result between jvp and vjp with the pdf / 1/Ns here

netket/netket/optimizer/qgt/qgt_jacobian_pytree_logic.py

Line 51 in 69331d8

res = tree_conj(_vjp(oks, _jvp(oks, v).conjugate()))

similarly to the onthefly implementation.

[codecov]

Codecov Report

All modified lines are covered by tests ✅

Comparison is base (45d334e) 83.74% compared to head (45b4586) 84.83%.
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1614      +/-   ##
==========================================
+ Coverage   83.74%   84.83%   +1.08%     
==========================================
  Files         262      262              
  Lines       14636    15140     +504     
  Branches     2210     2223      +13     
==========================================
+ Hits        12257    12844     +587     
+ Misses       1832     1761      -71     
+ Partials      547      535      -12     

Files	Coverage Δ
netket/jax/_jacobian/logic.py	88.00% <100.00%> (-1.37%)	⬇️
netket/optimizer/qgt/qgt_jacobian_dense.py	97.05% <ø> (+0.12%)	⬆️
netket/optimizer/qgt/qgt_jacobian_pytree.py	98.19% <ø> (+0.08%)	⬆️

... and 83 files with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[PhilipVinc]

@inailuig yes, you're right. I will do that in a followup PR removing this temporary flag... I'll try to keep this in mind.