Rc/0.1.6 - #1

- create() returned map_impl() directly, so a mapping failure on a segment
  this call had just created skipped the unlink that the fstat/ftruncate
  error paths perform; close() never unlinks, leaving the segment in
  /dev/shm until removed by hand.
- Route the mapping failure through cleanup_error(), which unlinks only
  when owns_shm_ is set.
- Add a POSIX regression test asserting that a failed create leaves nothing
  behind; verified it fails before the fix and passes after.
- Update platform notes and changelog.

- open() returned map_impl() directly, so a mapping failure left the fd open
  until the object was destroyed - inconsistent with the fstat() error path
  just above, and observable for a no-throw construction whose failed object
  stays alive.
- Add a regression test that counts descriptors while the failed object is
  still in scope; verified it fails before the fix and passes after.

- create() passed PAGE_READONLY to CreateFileMapping when a segment was
  created read_only, making the section read-only for every process forever;
  no later opener could map it for writing. The same code on POSIX opens
  O_RDWR and only mmaps PROT_READ, leaving the segment writable for others.
- Always create the section PAGE_READWRITE and let get_map_access() restrict
  this process's view, so read_only means the same thing on both platforms.
  Verified the creator's own view is still read-only (PAGE_READONLY / r--s).
- Add a test that a read-only creator does not lock out a read-write opener;
  verified it fails before the fix and passes after.
- Document the semantics in the API reference and platform notes.

- api_reference.md said "Always create (truncate if exists)", but only Linux
  resizes an existing segment: macOS resizes only when no other process holds
  it open, and Windows never resizes at all - neither growing nor shrinking.
  size() can therefore come back smaller than requested, and on Linux an
  existing segment can be shrunk while other processes have it mapped.
- Add a per-platform table and the size() check idiom to the API reference,
  broaden the macOS-only note in the platform notes, and add a README
  best-practice line.
- Pin the per-platform outcome with a test; macOS is left unasserted since it
  depends on whether the segment is held open elsewhere.

- unique_name() and shm_cleanup were copy-pasted verbatim across four test
  files; move them to tests/test_helpers.hpp in a named slick_shm_test
  namespace, with unique_name marked inline.
- Drop the now-unused <chrono> includes and the anonymous namespaces left
  empty; count_open_fds() stays local to test_error_handling.cpp.
- List the header in the test target so it appears in IDE project views.

- The view dropped its owned std::string for a borrowed const char*, which
  went stale whenever the source shared_memory was moved: small names live
  inside the string object, so a move left the pointer aimed at the moved-from
  object. Length-dependent and silent - long names survived, short ones did
  not, and a std::vector<shared_memory> reallocation broke it while satisfying
  the documented "source must outlive the view" rule.
- Store the name behind shared_memory::name() in address-stable storage
  (detail::stable_name) so a move transfers a pointer and leaves the
  characters put. sizeof(shared_memory_view) 56 -> 32 with noexcept copies,
  sizeof(shared_memory) 112 -> 88.
- Reject constructing a view from a temporary at compile time.
- Add tests/test_view.cpp; the class had no coverage. Verified the move,
  move-assign and vector-realloc cases fail without the stable storage.