Restore three-selection RO/RW design, drop eager world<->local flushes

Restores PR isaac-sim#5728's three-selection layout (`_trans_sel_ro`,
`_world_sel_rw`, `_local_sel_rw`) with asymmetric Fabric access flags
on `worldMatrix` and `localMatrix`.  Those flags are what protect the
user's write from being clobbered by Kit's per-tick
`IFabricHierarchy.update_world_xforms`:

* On `set_world_poses` (via `_world_sel_rw`, `localMatrix=RO`), Fabric
  does not recompute world from local -- the user's worldMatrix write
  survives until the renderer reads it.
* On `set_local_poses` (via `_local_sel_rw`, `worldMatrix=RO`), Fabric
  recomputes world from the new local on the next tick -- the renderer
  reads the correct world.

A single combined `worldMatrix=RW, localMatrix=RW` selection (the
recent design on this branch) removed that protection.  Fabric saw
both attributes as user-authored and fell back to the hierarchy's
canonical direction (local -> world), recomputing world from a stale
local and silently overwriting the user's world write.  That was the
failure mode behind the
`test_output_equal_to_usdcamera` regression and any other Camera + RTX
path that drives world poses through Fabric.

With the RO/RW protection back in place, the eager world<->local
flushes introduced by commits "fix: flush Fabric world matrices after
local writes" and the follow-up "set_world_poses eager local sync" are
no longer needed and are removed.  The `change_block` context manager
and its companion helpers existed only to batch those eager flushes;
with the flushes gone, the API has nothing to defer and is removed
from both `BaseFrameView` and `FabricFrameView`.

Class docstring now spells out the load-bearing role of the RO/RW
layout so a future refactor doesn't reintroduce the single-selection
shape.

Tests:

* Removed `test_set_local_*_updates_renderer_facing_fabric_world_matrix`
  (asserted an eager-update contract that the lazy design deliberately
  does not hold; correctness across the next render tick is provided
  by the RO/RW protection, not by an extra Warp kernel).
* Removed the four `test_change_block_*` tests; the API is gone.
* Inverted `test_interleaved_set_emits_no_warning` back to
  `test_interleaved_set_emits_warning`, restored `_dirty == LOCAL`
  assertions in `test_world_scales_roundtrip` and the symmetric
  `WORLD` assertion in `test_local_scales_roundtrip`, and updated
  `test_multi_view_per_view_dirty_isolation` to expect the lazy
  cross-view behavior.
* Adapted `test_prepare_for_reuse_detects_topology_change` and
  `test_fabric_rebuild_after_topology_change` to poll/rebuild all
  three selections.

Verified:

* `pytest test_ray_caster_camera.py::test_output_equal_to_usdcamera` passes.
* `pytest test_ray_caster_camera.py::test_output_equal_to_usd_camera_when_intrinsics_set` 4/4 pass.
* `pytest test_views_xform_prim_fabric.py` 71 passed, 3 skipped (cuda:1).
* `./isaaclab.sh -f` clean.

Net diff -198 lines.

Author: pv-nvidia

source/isaaclab/isaaclab/sim/views/base_frame_view.py

@@ -8,7 +8,6 @@
 from __future__ import annotations
 
 import abc
-import contextlib
 import warnings
 
 import warp as wp
@@ -39,37 +38,6 @@ def device(self) -> str:
         """Device where arrays are allocated (``"cpu"`` or ``"cuda:0"``)."""
         ...
 
-    @contextlib.contextmanager
-    def change_block(self, update_world_matrices: bool = True, update_local_matrices: bool = False):
-        """Batch multiple transform writes into one logical change.
-
-        Backends may use this context to defer expensive derived-state updates
-        until the outermost block exits.  The default implementation is a no-op
-        so callers can use it with every FrameView backend.
-
-        Args:
-            update_world_matrices: Whether derived world matrices should be
-                updated before the outermost block exits.  Backends may ignore
-                this option when they do not maintain separate cached world
-                matrices.
-            update_local_matrices: Whether derived local matrices should be
-                updated before the outermost block exits.  Backends may ignore
-                this option when they do not maintain separate cached local
-                matrices.
-        """
-        yield self
-
-    def changeBlock(
-        self,
-        updateWorldMatrices: bool = True,
-        updateLocalMatrices: bool = False,
-    ):
-        """CamelCase alias for :meth:`change_block`."""
-        return self.change_block(
-            update_world_matrices=updateWorldMatrices,
-            update_local_matrices=updateLocalMatrices,
-        )
-
     @abc.abstractmethod
     def get_world_poses(self, indices: wp.array | None = None) -> tuple[ProxyArray, ProxyArray]:
         """Get world-space positions and orientations for prims in the view.