[inplace_vector] WIP, allocator parameter

This might be the direction that P0843 takes after R10.
It has some advantages, but boy it's a lot more complicated.

README.md

@@ -101,7 +101,7 @@ Boost also provides all four adaptors; see [`boost::container::flat_set`](https:
 ```
 #include <sg14/inplace_vector.h>
 
-template<class T, size_t N>
+template<class T, size_t N, class A = allocator<T>>
 class sg14::inplace_vector;
 ```
 
@@ -113,6 +113,13 @@ and will likely be in C++26. The `sg14` version is portable back to C++17.
 
 Boost provides this container under the name [`boost::container::static_vector`](https://www.boost.org/doc/libs/1_83_0/doc/html/container/non_standard_containers.html#container.non_standard_containers.static_vector).
 
+#### Why the allocator parameter?
+
+Typical users of `inplace_vector` should simply pretend the allocator parameter isn't there.
+`inplace_vector` will never call `A::allocate` nor `A::deallocate`. But it will still respect
+`A::construct` (important to users of "scoped" allocators like `std::pmr`), and its iterators
+will be `A::pointer` (possibly important to users of Boost.Interprocess).
+
 #### Why not `std::erase{,_if}(inplace_vector, x)`?
 
 We shouldn't inject our own overloads into `namespace std`.