After moving an IfcGridAxis in Blender, the drawing annotation
was not tracking the axis to its new visual position. Two issues
were found and fixed:

1. generate_grid_axis_reference_points used the IFC AxisCurve
   geometry (via create_shape) with the grid object's matrix_world.
   After a save, the IFC AxisCurve is updated but the Blender mesh
   is not rebuilt, causing the two sources to diverge. The fix reads
   the axis object's Blender mesh vertices directly with
   axis_obj.matrix_world, which always matches what Blender renders.

2. When no Blender axis object exists, falls back to reading IFC
   geometry with the grid object's matrix_world (unchanged behavior).

Minor refactors: extracted matrices_differ variable in
sync_grid_axis_object_placement (export_ifc.py and drawing.py) and
extracted grid_placement variable in create_axis_curve.py for clarity.

Generated with the assistance of an AI coding tool.

When duplicating an IfcGridAxis one or more times before saving,
duplicates shared the same IfcPolyline as the source via shallow copy.
This caused two issues: (1) updating any one axis's AxisCurve during
export would destroy the shared curve, corrupting others; (2) duplicates
whose matrix_world checksum happened to match their current position were
skipped entirely by the is_moved guard, so their moved position was never
written to IFC.

Three fixes:

- geometry.py: call create_axis_curve immediately after copy_class for
  IfcGridAxis duplicates, so each new axis owns its AxisCurve from the
  moment of duplication rather than sharing the source's.

- create_axis_curve.py: only remove the old AxisCurve when its inverse
  count drops to zero, preventing destruction of curves still referenced
  by other axes.

- export_ifc.py: move the IfcGridAxis branch before the is_moved guard.
  Grid axes store position in AxisCurve geometry rather than
  ObjectPlacement, so is_moved is not a reliable gate. The internal
  matrices_differ check is the correct decision point, and
  record_object_position at the end keeps checksums in sync.

Generated with the assistance of an AI coding tool.

In IFC4X3, IfcGrid is a subtype of IfcPositioningElement, but
IfcGridAxis is not. The unlock handler only fetched
IfcPositioningElement instances, so grid axes were never
unlocked and remained immovable.

Fix: include IfcGridAxis in the element list for the IFC4X3
(non-IFC2X3/IFC4) branch of update_grid_is_locked.

- Add test_create_axis_curve.py with three unit tests: basic
  polyline creation, safe removal of an unshared existing curve,
  and preservation of a shared curve when only one referencing
  axis is updated (regression for the shallow-copy duplication bug)
- Add feature scenario "Export IFC - with duplicate-of-duplicate
  grid axis locations preserved" to project.feature, reproducing
  the case where duplicates of duplicates lost their positions on
  save/reload