`with_critical_section_mutex` improvements

[ngoldbaum]

Over in #6050 we fixed an issue with possible UB using with_critical_section_mutex2. However, there are additional issues:

[The] API allows possible UB by passing the same mutex twice and calling get_mut and then get (so there is a simultaneous mutable and immutable reference).

I think we can fix this by changing the API of EnteredCriticalSection to be a new EnteredCriticalSectionUnsafe struct:

#[cfg(all(Py_3_14, not(Py_LIMITED_API)))]
pub struct EnteredCriticalSectionUnsafe<'a, T>(&'a UnsafeCell<T>);

#[cfg(all(Py_3_14, not(Py_LIMITED_API)))]
impl<T> EnteredCriticalSectionUnsafe<'_, T> {
    pub unsafe fn get(&self) -> *const UnsafeCell {...}
}

That way if anyone needs unsafe aliased access to the data, they have to explicitly do it via the UnsafeCell API.

I think this is only necessary for the _mutex variants and we shouldn't need to change the API of with_critical_section or with_critical_section2, since those don't allow locking arbitrary data.

Credit to @epilys for suggesting this to me on IRC.

Also as pointed out on the PR review by @davidhewitt, we can also probably use MaybeUninit instead of std::mem::zeroed() when initializaing CsGuard and Cs2Gaurd internally:

pyo3/src/sync/critical_section.rs

Line 200 in 072d544

let mut guard = CSGuard(unsafe { core::mem::zeroed() });

[Person-93]

Maybe the function passed to with_critical_section_mutex2 should take an Option<EnteredCriticalSection> as its second param and if the two mutexes are the same pointer, it can be set to None.