limit answers 1 for z/abs(z) whatever the phase of z #1186

limit answers 1 for z / abs(z) and z ^ 2 / abs(z) ^ 2 whatever the phase of z, and the
answer is wrong wherever that phase is not real and positive.

limit((i/x)/abs(i/x), x, +oo)                  ->  1     true value: i
limit((i/x)^2/abs(i/x)^2, x, +oo)              ->  1     true value: -1
limit(((1+i)/x)^2/abs((1+i)/x)^2, x, +oo)      ->  1     true value: i

Checking the first by hand: for real x → +oo, i/x is a positive imaginary number and
abs(i/x) = 1/x, so the quotient is i at every x and its limit is i. The second: (i/x)^2 is
-1/x^2 while abs(i/x)^2 is 1/x^2, so the quotient is -1 everywhere. The third: (1+i)^2 = 2i
and abs(1+i)^2 = 2, so the quotient is i everywhere. None of the three is a limit that needs
taking — the expression is constant in x.

Simplify is not at fault

"x ^ 2 / abs(x) ^ 2".Simplify()   ->  (x / abs(x)) ^ 2        at x = i: -1, correctly
"abs(x) ^ 2".Simplify()           ->  abs(x) ^ 2              at x = i:  1, correctly

So the simplifier keeps the phase and evaluates correctly. It is the limit machinery that discards
it — Absf's limit is computed as abs(<limit of the argument>), which is true for a real path and
loses the argument's direction otherwise.

The same cause, without an i in the input

limit(x ^ 2 / abs(x) ^ 2, x, 0)   ->  1

The default codomain is Complex, and along the imaginary axis x = it the quotient is
-t^2 / t^2 = -1, so no complex limit exists and 1 is a wrong answer rather than an
unqualified one. Under Codomain = Real the same query answers 1, which is right — so the defect
is specifically that the complex reading is being answered with the real one.

limit(x / abs(x), x, 0) already answers NaN, which is correct, so the machinery is not uniformly
wrong — it is wrong where the quotient's phase is constant and non-real, which is exactly where it
looks like it can cancel.

Not the same as #720

#720 asks for ApproachFrom to be able to
name a direction in the complex plane. This needs no direction named: every case above is reachable
today with the ordinary two-argument limit, and three of them return a wrong value rather than
declining. Worth fixing first and separately — a direction parameter on a machinery that mishandles
abs off the real line would inherit this.

Found while measuring #720 rather than from a report, so no user is known to have hit it.