some fixes to method ambiguity reflection

[JeffBezanson]

Putting #16125 through its paces, I discovered the following case where methods returned duplicates:

julia> amb_2(::Int, y) = 1

julia> amb_2(x, ::Int) = 2

julia> amb_2(::Int8, y) = 3

julia> methods(amb_2)
#4 methods for generic function "amb_2":
amb_2(::Int64, y) at REPL[1]:1
amb_2(x, ::Int64) at REPL[2]:1
amb_2(::Int8, y) at REPL[3]:1
amb_2(x, ::Int64) at REPL[2]:1

I then decided that the default behavior of methods(f, t) should be to return only methods that could actually be called, so ambiguous matches should be excluded. This PR implements that and fully avoids returning duplicate entries.

I also found this evil case:

amb_3(::Int8, ::Int8) = 1
amb_3(::Int16, ::Int16) = 2
amb_3(::Integer, ::Integer) = 3
amb_3(::Integer, x) = 4
amb_3(x, ::Integer) = 5

Here, the last two methods appear ambiguous but aren't, because their intersection is fully covered by prior methods. isambiguous did not handle this.

I have added tests for these cases.

[vtjnash]

+1

alternatively, I've been wondering if we can use the work done here to create a fast-path for invalidate_conflicting where we only test lambdas that were derived from methods that check_ambiguous deemed to have an intersection

[JeffBezanson]

This was motivated by the following example:

f(::Int8, x)
f(::Integer, x)
f(x, ::Int)

Here we fail to detect that Int8, Int is ambiguous.

Yes, that sounds promising.

[timholy]

Wow, thanks for noticing those problems and fixing them.