Simplify inverts a quotient whose denominator has a negative factor

[Rafael-SOWNet]

Simplify returned the reciprocal of the right answer for a quotient whose denominator carried a negative factor.

"(1/x) / (-1 - 1/x)".ToEntity().Simplify()   =>   -(1 + x)

At x = 1 the expression is -1/2 and that answer is -2.

The cause is in Patterns.NumericNeatRules. Four of the rules there take a negative constant out of a product and multiply the rest by it, which is right. The two that take it out of a denominator were written the same way:

Divf(var any2, Mulf(Real { IsNegative: true } num, var any1)) => -((-num) * (any1 / any2)),

but a / (-b * c) is -(a / (b * c)), not -(b * (c / a)). Simplify keeps whichever of its candidates is shortest, so the inverted form won wherever it was smaller than the correct one -- which is wherever the numerator is 1, hence the shape above.

Found from a limit rather than by hand: l'Hopital's rule simplifies the quotient of derivatives it builds, and (1/x) / (-1 - 1/x) is the shape of one. lim x->0+ (tan(x) * ln(x)) came back NaN because of it.

Filing for the record -- the fix is in #714 with five regression cases, each checking the value at a point rather than the printed form.

[Rafael-SOWNet]

Fixed by #714, which is merged. Re-measured on master at 7a38a8a:

	before	now
((1/x) / (-1 - 1/x)).Simplify()	-(1 + x), which is -2 at x = 1	unchanged, and -1/2 at x = 1
(1 / (-2 * x)).Simplify()	inverted	-1/2 / x
lim x->0+ tan(x) * ln(x)	NaN	0

The last line is the one worth recording: this was found from a limit rather than by hand, because l'Hopital's rule simplifies the quotient of derivatives it builds and (1/x) / (-1 - 1/x) is the shape of one. That limit is now answered.

Five regression cases guard it in SimplificationRegressionTest.ANegativeFactorUnderTheLineStaysUnderIt, each checking the value at a point rather than the printed form — an inverted answer differs from the original by a factor, and subtracting the two can simplify to something that only looks like zero if the same wrong rule fires on the difference. All five pass.

One thing the fix does not do, noted separately since it is not this defect: (1/x) / (-1 - 1/x) now comes back unchanged rather than reduced to -1/(1 + x). That is correct but not simplified — the reduction it used to perform was the buggy one, and no other route reaches the right one.