Fix handling of allow_bottom_tparams in isambiguous

[timholy]

With master, we have a lot of ambiguities, but a subset of these involve method signatures that are ambiguous only because of the presence of Union{} as a type-parameter in the intersection:

julia> foo(x::Complex{<:Integer}) = 1
foo (generic function with 1 method)

julia> foo(x::Complex{<:Rational}) = 2
foo (generic function with 2 methods)

julia> m1, m2 = collect(methods(foo))
2-element Array{Any,1}:
 foo(x::Complex{#s1} where #s1<:Integer) in Main at REPL[3]:1 
 foo(x::Complex{#s1} where #s1<:Rational) in Main at REPL[4]:1

julia> typeintersect(m1.sig, m2.sig)
Tuple{#foo,Complex{Union{}}}

julia> Base.isambiguous(m1, m2)
true

This isn't the kind of intersection we should usually be worried about, so in e101000 a strategy was developed to exclude these kinds of intersections. However, I found that it doesn't correctly catch all cases, including the one above:

julia> Base.isambiguous(m1, m2, false)
true

With this PR, I believe I've fixed the detection of Union{}. It has a pretty dramatic effect on our ambiguity list.

Master:

julia> length(Base.Test.detect_ambiguities(Base,Core,allow_bottom=true))
Skipping Base.<|
Skipping Base.Parallel
194

julia> length(Base.Test.detect_ambiguities(Base,Core,allow_bottom=false))
Skipping Base.<|
Skipping Base.Parallel
111

This PR:

julia> length(Base.Test.detect_ambiguities(Base,Core,allow_bottom=true))
Skipping Base.<|
Skipping Base.Parallel
194

julia> length(Base.Test.detect_ambiguities(Base,Core,allow_bottom=false))
Skipping Base.<|
Skipping Base.Parallel
11

11 seems rather more manageable (the "missing" digit is not a copy/paste error). Suddenly the steps that @Sacha0 has been taking (#20970, #20971) don't seem so small anymore: those should fix 3 of the 11. Of the rest, it's just 4 functions: cat (4 ambiguous pairs), checkbounds_indices (2 ambiguous pairs), colon (1 ambiguous pair), and a constructor method for Tuple (1 ambiguous pair).

I do find myself wishing we could change the keyword argument to consistently have the same name, and I find accept_bottom_tparams clearer than allow_bottom_tparams (and perhaps flipping the sense, exclude_bottom_tparams, to be the clearest of all). But at this point, I don't think we can do that for 0.6. I did add some docstrings.

Fixes #20246

[tkelman]

looks like this could be a doctest with minor changes

[Sacha0]

Insofar as I am qualified to review (not far), this looks great!

[Sacha0]

(#20970, #20971, #20986, #20988, #20989, and #20990 should together resolve all relevant ambiguities. Best!)

[timholy]

@Sacha0, when your pile of PRs merges, please do us the honor of turning that test back on. Thanks for jump-starting the effort to resolve these ambiguities.

[Sacha0]

exclude_bottom_tparams, to be the clearest of all

+1

[tkelman]

unrelated aside, but I think we should gradually aim to get away from using square brackets to denote optional arguments and uniformly use the julia syntax convention for them in docstrings