Preserve PyPose LieTensor semantics in TrackingTensor

[zitongzhan]

Summary

This PR makes TrackingTensor work transparently with PyPose LieTensor inputs so tracked SE(3) parameters keep their LieTensor API instead of degrading to plain tensors.

The main user-facing effect is that code like self.nodes = nn.Parameter(TrackingTensor(nodes)) now preserves pp.SE3 behavior through indexing, so PGO code can use:

  (poses.Inv() @ node1.Inv() @ node2).Log().tensor()

instead of repeatedly re-wrapping everything with pp.SE3(...).

What Changed

• Made TrackingTensor(...) polymorphic for pp.LieTensor inputs.
• Added an internal Lie-aware tracking subtype that preserves ltype across:
  • nn.Parameter(...)
  • tensor indexing
  • torch.cat(..., dim=0)
• Kept the plain tensor tracking path unchanged for non-LieTensor inputs.
• Updated the Jacobian graph path to preserve LieTensor metadata during torch.vmap(jacrev(...)) via pp.retain_ltype().
• Kept Jacobian trace shapes unchanged during backward accumulation and only applied trim_SE3_grad when materializing the final Jacobian.
• Added explicit jactrace cleanup before and after Jacobian construction to avoid stale trace state leaking across calls.
• Simplified the PGO example to use native LieTensor operations directly.
• Fixed a test typo in the BAL fixed-camera helper (model.points_3d vs model.points).

Why

Previously, wrapping a pp.LieTensor in TrackingTensor stripped its LieTensor type and ltype, which forced downstream code to re-cast with pp.SE3(...) before every Lie operation.

This PR keeps the sparse-tracing behavior while preserving the LieTensor API, which makes the PGO path cleaner and less error-prone.

• tracked LieTensor indexing/cat preserving ltype
• PGO-style SE(3) residuals working without explicit pp.SE3(...) rewraps