This is a bit of hack in case some person does not pass __file__ to taskmaker and expects their jobname to come out exactly as they put it in. Let us solve this when somebody complains.

I wrote a custom mt19937 implementation to replace it.
Using the available libraries (stl, intel_mkl) in production
code is not feasible because significant parts of them cannot
be inlined. Our implementation is (like the old one) header-only
and inlined completely.

It is also gives identical random sequences for a given (internal)
seed as the old MersenneTwister.h and std::mt19937. As it is checkpoint
compatible with MersenneTwister.h, it should not affect existing
runs.

otherwise slurm will kill the job after a small timeout

on claix18, batch jobs should use whatever is set in $MPIEXEC (srun at the moment)
using mpirun instead works, but it seems that it leads to some bad pinning of ranks which
reduces CPU efficiency to just 2%!

jobs

instead of $(pwd)/JOBNAME.data, the job director is now the directory containing the parameter file.
This will not change the behavior of the loadl script, which changes directory and creates the parameters file at
$(pwd)/JOBNAME.data/parameters.json

However, when you used loadleveller without loadl, there were subtle complications when you do not get the working directory hopping correct.
The new behavior is more predictable and helps with that. One slightly ugly artifact is that the .results.json file gets created in “the parent directory of the job directory” which can be really messy if you don’t follow the same convention loadl uses. But I’ll leave it like this for
the sake of simplicity.

For good measure, I have extended the use of std::filesystem::path in the code.

* eliminates a source of bugs (relying on state that was not set in the mc constructor)
* allows speeding up your merges: mine were slow because of all the things that happen in the constructor

However you may need to rearrange your code to calculate all the things you need for your evalables in register_evalables.

it doesn’t seem to be faster than internal_mersenne still. but maybe in an actually simulation for some cases it is?

and disable python backtraces because it mostly gets in the way of the user

at the cost of some i/o cost (one copy of the meas file at each checkpointing),
we can now assure that the results of the simulation are always in a valid state,
even if the program crashes while writing them out. At least in theory. Let’s see
how it works.

We also changed to the std::filesystem API… meaning there could be bugs…
and loadleveller probably runs on Windows now!

Changed default argument in observable.h: one can use intel compiler now

See merge request !1

up to now seeds were not saved to the dump file. Since the state of the
RNG was saved properly and the seed is only used in the initialization,
this likely did not affect the results.