Update backend: x^y instead of abs(x)^y, and high-precision constants

[MilesCranmer]

This updates the backend to v0.12, which includes a couple of changes:

1. Use functions returning NaN on branch cuts instead of abs (issue #109) SymbolicRegression.jl#123 by @johanbluecreek, which changes operators from, e.g., abs(x)^y to x^y. Expressions using these operators will now simply avoid cases where their output is NaN. This avoids letting the search exploit weird functional forms which one might wish to not use.
2. Generalize expressions to have arbitrary constant types SymbolicRegression.jl#119, which uses arbitrary-precision for constants. Now, when you use PySRRegressor(precision=64), the constants will also be 64-bit precision. As a side effect, this also increases the speed of evaluation, since there is type stability.
3. Fix for when the dataset has zero variance. This allows you to fit datasets with a single row: useful if you want to find a symbolic approximation to a number (Ramanujan-style), like described on express a number as a function of given constants SymbolicRegression.jl#127.
4. This also sets the default parsimony to 0.0. I felt like it makes a lot of assumptions about the dataset scale to set any value for the default parsimony, so it's better to just leave the default turned off. The search will still find simple expressions; but it will not be as biased towards them, especially when the baseline loss is large.

[MilesCranmer]

For example, here is how you can find the result of pi/sqrt(2) with PySR:

import math
from pysr import PySRRegressor

model = PySRRegressor(
    binary_operators="+ * / -".split(" "),
    unary_operators=["sqrt"],
    precision=64,
    complexity_of_constants=100,
    niterations=1000,
    weight_mutate_constant=0.0,
    should_optimize_constants=False,
    parsimony=0.001,
)

model.fit(
    [[2, math.pi, math.e]], [2.22144146908], variable_names=["_2", "_pi", "_e"]
)

At the end, you should be able to run:

model.sympy()
# _pi/sqrt(_2)

[MilesCranmer]

Finding approximations to pi :)

import math
from pysr import PySRRegressor

model = PySRRegressor(
    binary_operators="+ * / -".split(" "),
    unary_operators=["sqrt", "square"],
    niterations=1000,
    # High precision:
    precision=64,
    # Favor simpler expressions:
    parsimony=0.003,
    # Disable constants:
    complexity_of_constants=100,
    # Prevent redundant computation:
    weight_mutate_constant=0.0,
    should_optimize_constants=False,
)

model.fit(
    [[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]],
    [math.pi],
    variable_names=["_1", "_2", "_3", "_4", "_5", "_6", "_7", "_8", "_9", "_10"],
)