Propagate py::multiple_inheritance to all children

[virtuald]

Description

This builds on the tests split from #3665 into #3649, which should probably be merged first.

This started with a comment that @Skylion007 made on my tests for #3635, and I realized the fix was straightforward. Previously, you had to mark any child of a class that participated in multiple inheritance with py::multiple_inheritance, even if the child derived from a single base. Now, pybind11 can autodetect this in most cases.

I'm not confident this is quite the right fix. It's not clear to me what the difference between simple_type and simple_ancestors are, and I wonder if the purpose of simple_type has evolved over time as it's only used in one place.

A different fix might be to change type_caster_base's simple_type check to simple_ancestors, but then simple type wouldn't be used anywhere?

... actually, I just realized this change makes simple_type and simple_ancestors identical? Simple ancestors is only used for registering/deregistering instances.

It's late here, I'll sleep on it. There must be a subtle side effect of simple_type that I'm just missing and that the tests don't currently cover.

Suggested changelog entry:

* ``py::multiple_inheritance`` is now only needed when C++ bases are hidden from pybind11

[Skylion007]

According to the documentation and our testing for the TypeInfo type, a type with non simple_ancestors can never be a simple type. We should add an assert somewhere that verifies this:

Suggested change

	// a simple type can never have non-simple ancestors
	assert(!tinfo->simple_type||tinfo->simple_ancestors));

[rwgk]

@jagerman for comment, who I believe was the main author of the multiple inheritance code.

[rwgk]

I'll try to get this through the global testing asap (hopefully the 12:00 batch, in ~2 hours).

[rwgk]

This PR is now in the global testing pipeline (12:00 batch), after first testing manually with ASAN.

[rwgk]

The global testing went fine. Results are a bit noisy again, but after looking through I'm convinced we're good to go ahead.

[rwgk]

This is also on the smart_holder branch now. Thanks again!