Fixup hash(::Real) broken by #49996

[LilithHafner]

I mishandled nonzero pow returned by decompose(::BigFloat).

Fixes #50065 and adds a test

[vtjnash]

Please use meaningful titles for commits and PRs and not #'s

[oscardssmith]

add 1//1 and big(1) for good measure.

[LilithHafner]

I think it makes sense to add BigFloat and BigInt to the massive cross-product at the top instead.

[oscardssmith]

well I'm sure glad I suggested testing BigInt :)

[LilithHafner]

My justification for changing the collisions bound and comment describing it is that the comment is misleading.

julia> collides = [];

julia> for T = types, S = types, x = vals
           a = coerce(T, x)
           b = coerce(S, x)
           eq = hash(a) == hash(b)
           #println("$(typeof(a)) $a")
           #println("$(typeof(b)) $b")
           if isequal(a, b)
               @test eq
           elseif eq
               push!(collides, (a,b)) 
           end
       end
       # each pair of types has one collision for these values

julia> using StatsBase; countmap(typeof.(collides))
Dict{DataType, Int64} with 40 entries:
  Tuple{UInt64, Int8}                      => 2
  Tuple{UInt64, Rational{Int16}}           => 2
  Tuple{Rational{Int8}, Rational{UInt64}}  => 2
  Tuple{Float32, Rational{UInt64}}         => 9
  Tuple{Float64, Rational{UInt64}}         => 14
  ⋮                                        => ⋮

On average, each pair of types has 0.947 collisions. Adding BigFloat and BigInt increases this average to 1.076. The new bound is exactly the number of current collisions.

[LilithHafner]

New tests on this PR unearthed a hashing bug for BigInt: #50075. I'm going to skip BigInt tests for now.

[oscardssmith]

Sounds good. #50076 should fix that part of this.

[LilithHafner]

Test failure looks unrelated