Loosen rebuild-after-topology tolerance to 1e-5

The chained set_world_poses -> get_world_poses round-trip in
test_fabric_rebuild_after_topology_change goes through Warp's float32
SRT compose/decompose, which accumulates a few ULP of drift.

At the test's position magnitudes (~4-6), one float32 ULP is
~4.77e-7, so the prior atol of 1e-7 demanded sub-ULP agreement and
was sensitive to GPU/codegen variation -- it passed locally on the
A6000 but flaked in CI.

1e-5 corresponds to roughly 20 ULP at those magnitudes: tight enough
to catch any real bug (a wrong index or stale read would be at least
~1e-3 off given the test setup) and consistent with the shared
contract harness in frame_view_contract_utils.py, which already
documents and uses ATOL = 1e-5 for compose/decompose-through-float32
checks.

Author: pv-nvidia

source/isaaclab_physx/test/sim/test_views_xform_prim_fabric.py

@@ -209,7 +209,8 @@ def test_fabric_rebuild_after_topology_change(device, view_factory):
     ret_pos, _ = view.get_world_poses()
     pos_torch = wp.to_torch(ret_pos)
     expected = torch.tensor([[4.0, 5.0, 6.0], [4.0, 5.0, 6.0]], device=device)
-    assert torch.allclose(pos_torch, expected, atol=1e-7), f"Read after rebuild failed on {device}: {pos_torch}"
+    # 1e-5 ≈ 20 ULP at magnitudes ~4-6; absorbs float32 SRT compose/decompose drift.
+    assert torch.allclose(pos_torch, expected, atol=1e-5), f"Read after rebuild failed on {device}: {pos_torch}"
 
 
 @pytest.mark.parametrize("device", ["cuda:0"])