Switch Julia interface to PyJulia

[MilesCranmer]

This switches the entirety of the Julia/Python interface to use PyJulia rather than a subprocess call to a generated script. Multithreading and distributed computing both work!

This brings several advantages:

• Huge speedups in startup time on second pysr( ) calls in the same python call
• Simplified error debugging on the Julia side
• Greatly simplified codebase with no huge chunks of code generation
• Possibility to checkpoint the symbolic regression search, and restart later

The disadvantage is that this will put PySR at the mercy of PyJulia and any bugs that it experiences. However, I think in the longterm this is still much better than running a script of generating code.

Installation with this new interface is now:

pip install pysr
python -c 'import pysr; pysr.install()'

This will install PyCall.jl (required by PyJulia) as well as the packages defined in SymbolicRegression.jl.

Will initially be implemented in v0.7.0a1 for a while until things stabilize.

TODO:

• Update documentation to describe PyJulia install.
• Update documentation to mention that repeat calls will have a quicker startup time.
• Check that error is correct when PyJulia not installed if pysr.install() has not been run.
• Check if there are any performance hits using PyJulia vs Julia.
• Make it so we can KeyboardInterrupt the PyJulia execution of EquationSearch. See KeyboardInterrupt support / SIGINT handler JuliaPy/pyjulia#211.
• This does not work when python is statically linked. In this situation, should we use compiled_modules=False? Or enfoce that the user has to get this working?
• Is there anything I can do to get conda working with PyCall.jl?
• Test conda with the CI.

[MilesCranmer]

To get KeyboardInterrupt to work, I'm going to launch EquationSearch in a multiprocessing.Process. Upon KeyboardInterrupt, a signal will be given to a temp file. The actual SymbolicRegression.jl EquationSearch will check for this signal during the search loop. It will then manually exit if this signal is set to 1.

Update: instead, I will have SymbolicRegression.jl itself read stdin constantly, and if "q" is entered, it will quit and return the current set of populations and hall of fame. This works!

[MilesCranmer]

So it seems like PySR actually speeds up with PyJulia versus the old method of launching a script.

Measurements before the change:

[64700., 67500., 68800., 68900., 69300., 69800., 68400., 73000.,
       74600., 75100., 75500., 75400., 57700., 59700., 60600., 61300.,
       61900., 62500., 62900.]

(average 67242)
Measurements after the change:

[ 87800.,  87900.,  88000.,  87700.,  87200.,  90900.,  90800.,
        89500.,  89000.,  88700., 121000., 121000., 121000.]

(average 96192).

These measurements are very small jobs but this at least shows it doesn't drastically hurt performance, so I think we are good on that front.