test: math: test floating-point behavior when mapping zero to zero

[nsajko]

The idea for this addition to the test suite originates here, where stevengj suggested to do this test for cosc specifically:

• improve cosc(::Float32) and cosc(::Float64) accuracy #59087 (comment)

This PR instead tests all applicable functions I could find.

Also fix the sign of cosc(::BigFloat) for zero input, thanks to stevengj.

[nsajko]

cosc(::BigFloat) is broken currently at the origin when testing it like this. Thus this PR doesn't test BigFloat, that can be left for a future PR.

[stevengj]

cosc(::BigFloat) is broken currently at the origin when testing it like this.

Can be fixed by taking the line s = (term = -(π*x))/3 in the cosc(::Number) implementation and adding

iszero(s) && return s

immediately afterwards.

(Otherwise it computes one extra term in the Taylor series, which messes up the sign of zero because -0.0 + 0.0 === 0.0.)

[stevengj]

Should it also test binary operators, like Base.Fix1(*, one(typ)), Base.Fix2(*, one(typ)), Base.Fix2(/, one(typ))?

[stevengj]

This looks questionable:

julia> div(-0.0, 1.0)
0.0

[stevengj]

Should be good to merge.

In principle, we could backport the cosc(big"0.0") sign fix? Maybe too minor to bother with.