Turn a negated difference round wherever it sits (#882)

[Rafael-SOWNet]

Closes #882.

-(a - b) became b - a for a whole expression and not for the same expression inside another node,
so what a caller got depended on where the subexpression sat:

input	was	is
-(5 - sqrt(-11))	sqrt(-11) - 5	unchanged
-(5 - sqrt(-11)) + y	y - (5 - sqrt(-11))	sqrt(-11) - 5 + y
2 ^ (-(5 - sqrt(-11)))	left as written	2 ^ (sqrt(-11) - 5)
sgn(-(5 - sqrt(-11)))	left as written	sgn(sqrt(-11) - 5)
[[-(5 - sqrt(-11)), 1]]	left as written	[[sqrt(-11) - 5, 1]]
Expand([[(x+1)^2, 1]])	[[(x + 1) ^ 2, 1]]	[[1 + 2 * x + x ^ 2, 1]]

Every one of these was already the right number, written the long way round.

Why it only worked at the root

The step came from Expand, which the simplifier offers as a candidate for the root expression
only, and which does not descend into an exponent, a function's argument or a matrix. So it is written
as a rule instead: a unary minus parses as (-1) * x, so the shape is (-1) * (a - b), and turning it
round drops the multiplication and the negative constant together — five nodes for three, and the
complexity criteria charge four more for a negative real. It lands in InvertNegativeMultipliers,
whose other arm normalises 1 + (-x) into 1 - x: the same family, and already inside the bundle
SimplifyChildren applies everywhere. A rule walks the tree, which is the whole of the fix.

Expand of a matrix is the same story from the other end. It reads its argument as a sum, and a matrix
is not one, so a matrix left through the "if one is too complicated, return the current one" exit and
came back as it arrived. It now expands entry by entry. Factorize and Differentiate both descended
already — they are built out of rewrite rules — so Expand was the odd one out rather than matrices
being held back deliberately.

Where a caller meets it

EquationSystem.Solve returns a Matrix, so a solved system's entries were left in whatever form the
solver built them in:

MathS.Equations("x2 + y", "y - x - 3").Solve("x", "y").Simplify()

was:  [[-(-(5 - sqrt(-11)) / 2 + 3), (5 - sqrt(-11)) / 2], [-(-(5 + sqrt(-11)) / 2 + 3), (5 + sqrt(-11)) / 2]]
is:   [[-1/2 + -1/2 * sqrt(-11), 5/2 + -1/2 * sqrt(-11)], [1/2 * sqrt(-11) + -1/2, 1/2 * sqrt(-11) + 5/2]]

Two corrections to the issue's own analysis

Both would mislead whoever read it next, so they are worth stating rather than quietly fixing.

The sum case's candidate is generated. alternate::-(5 - sqrt(-11)) + y lists sqrt(-11) - 5 + y
second. It lost because the two forms tie: y - (5 - sqrt(-11)) and sqrt(-11) - 5 + y both
rate 22, and a tie is settled by whichever candidate was generated first — the accident the
RationaliseDenominator comment in that same file already warns about. Only the power, sgn and
matrix cases are the "never generated" kind the issue describes.

The complexity table in the issue compared the wrong pair. It read the input (9) against the good
candidate (7) and concluded the good form would win. The chosen output is 7 as well — Simplify did
improve 9 → 7, then picked one of two equally-rated 7s. And Complexity is not what ranks candidates
in any case; SimplifiedRate is, which weights a negative real at 4 and a negative power at 8. I have
added a rate:: probe alongside complexity:: in the workspace harness so this confusion is one
command away from being caught, with the mistake named in its comment.

The rule fixes all four cases including the sum, because b - a is strictly better than either
candidate and so wins the tie outright rather than depending on order.

Measured

The new tests fail 7 of 246 against master and pass 246 of 246 here, checked by restoring
master's two files underneath them.

Suite 6359 passed / 0 failed, F# wrapper 130 passed. casbench 117/119 with 0 wrong, rootcheck 596/596,
simpsweep 10463/10463 — the one that matters most here, since this changes the form of a great many
expressions and simpsweep is what checks a changed form still has the old value. propcheck 1340 checks
with 0 failures, crashcheck 1652 cases with 0 crashes, boundcheck unchanged at 2 disagreements.

Cut from master at fbf1dd77.