The poisoned_hash_helper header was relying on an implicit forward
declaration of std::hash located in <type_traits>. When we improve
the modularization of the library, that causes issues, in addition
to being a fundamentally non-portable assumption in the test suite.

It turns out that the reason for relying on a forward declaration
is to be able to test that std::hash is *not* provided if we don't
include any header that provides it. But testing that is actually
both non-portable and not really useful.

Indeed, what harm does it make if additional headers provide std::hash
specializations? That would certainly be conforming -- the Standard
never requires an implementation to avoid providing a declaration when
a given header is included, instead it mandates what *must* be provided
for sure. In that spirit, it would be conforming for e.g. `<cstddef>`
to define the hash specializations if that was our desire. I also don't
read https://wg21.link/P0513R0 as going against that statement. Hence,
this patch just removes that test which doesn't carry its weight.

Fixes llvm#56938