add some type info to Base to avoid excess recursion in inference

[JeffBezanson]

Fixes #33336.

This addresses some commonly-occurring cases where having too little type info makes inference see a lot of recursion in Base that is not actually possible.

Of course, this is not satisfying, since these are empirically determined (*). @vtjnash & I would also like to make inference's cycle handling smarter, but that will be a longer project.

Needs review from @mbauman : Are the dataid arguments to _isdisjoint indeed always UInt? And is it safe to assume that the length of an index in broadcast is an Integer?

(*) Methodology is as follows:
Apply this patch:

--- a/base/compiler/inferencestate.jl
+++ b/base/compiler/inferencestate.jl
@@ -245,6 +245,10 @@ function add_mt_backedge!(mt::Core.MethodTable, @nospecialize(typ), caller::Infe
 end
 
 function poison_callstack(infstate::InferenceState, topmost::InferenceState, poison_topmost::Bool)
+    if trace_inf[1]
+        print_callstack(infstate)
+        println()
+    end

(you will also need to add const trace_inf = [false] somewhere. Set Core.Compiler.trace_inf[]=true just before running the problematic call. In the output, look for calls from simple "innocent" Base functions that lead into package code or big nests of code like broadcast, that they should not be able to touch. For example:

#broadcastMD#59(Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}}, typeof(MDDatasets.broadcastMD), MDDatasets.CastType2{Number, 1, Number, 2}, typeof(Base.:(>)), Int64, MDDatasets.DataHR{T2} where T2)  [limited]
broadcastMD(MDDatasets.CastType2{Number, 1, Number, 2}, typeof(Base.:(>)), Int64, MDDatasets.DataHR{T2}) where {T2}  [limited]
>(Int64, MDDatasets.DataMD)  [limited]
fill_to_length(Tuple{Integer, Vararg{Any, N} where N}, Int64, Base.Val{_A} where _A)  [limited]
(::Type{Base.IteratorsMD.CartesianIndex{_A}})(Tuple{Integer, Vararg{Any, N} where N})  [limited]

There, fill_to_length, a low-level tuple utility, led to inference of >(Int64, MDDatasets.DataMD) which then calls broadcast, etc. That can be patched up by declaring the second argument to > in fill_to_length as Int (which it will always be, but inference can't tell).

[mbauman]

Approve on both counts; data_ids only ever returns UInt (and is documented to do such), and I would be very surprised if broadcast could work with non-Integer lengths (and length is documented to only return Integers).