I think it would be clearer if these were a bunch of separate ifs instead of this long chained thing (if every assignment fit on one line, that would also be clearer.)

That'll be much simpler to do after #10680 is merged and we later get to remove the old version of the API. Then the only functions we'll have to load are Tk_FindPhoto and Tk_PhotoPutBlock_NoComposite, which also means that the Windows and non-Windows versions will only need to differ by how they call _dfunc.
Not convinced it's worth changing for now.

Just to be clear, the endgame is to just do (e.g.)

void *_dfunc(void *lib_handle, const char *func_name);

int _func_loader(void *lib)
{
    // Fill global function pointers from dynamic lib.
    // Return number of successfully loaded functions.
    #define LOAD_TK(name) tk::name = Tk_##name##_t(_dfunc(lib, "Tk_" #name))
    return bool(LOAD_TK(FindPhoto)) + bool(LOAD_TK(PhotoPutBlock_NoComposite));
    #undef LOAD_TK
}

followed by OS-specific versions of _dfunc.

#10680 is merged now.

But the old API is still around (the simplification will only come after matplotlib.backends.tkagg is completely removed).

What error is set here instead?

good catch... fixed.

Actually no, there is indeed an error being set by _dfunc()

void *_dfunc(void *lib_handle, const char *func_name)
{
    // Load function `func_name` from `lib_handle`.
    // Set Python exception if we can't find `func_name` in `lib_handle`.
    // Returns function pointer or NULL if not present.

    void* func;
    // Reset errors.
    dlerror();
    func = dlsym(lib_handle, func_name);
    if (func == NULL) {
        PyErr_SetString(PyExc_RuntimeError, dlerror());  // here
    }
    return func;
}

(and similarly on Windows).

Left the original version as is.

Or here?

Same split up here, probably.