Fix thread safety in _elementtree.c

[Sebcio03]

Bug description:

_elementtree.c declares Py_MOD_GIL_NOT_USED but has several data races in the free-threaded build (audited for #116738):

• PyDict_GetItemWithError in the expat entity handler returns a borrowed ref that another thread can invalidate by modifying the entity dict
  concurrently
• clear_extra / element_resize modify self->extra and the children array without any lock — concurrent append() + clear() on the same element
  can crash
• element_get_text, element_get_tail, element_get_attrib return borrowed refs into the struct with no lock held by the caller — a concurrent write
  can free the pointed-to object before it's used
• All four property getters/setters (tag, text, tail, attrib) access struct fields with no critical section
• __copy__ snapshots struct fields without a lock
• __deepcopy__ has the same problem and additionally calls PyDict_Next on attrib without a critical section on the dict
• treebuilder_handle_end reads self->stack and decrements self->index without a lock

Fix and tests in #145569.

[da-woods]

pycore_critical_section.h was not included — all subsequent critical-section macros depend on it.

That at least is nonsense. The public critical section interfaces are in Python.h