fix: normalize ICP2d's returned transform to exact orthonormality

[petercorke]

Summary

• `ICP2d`'s final `T` accumulates ordinary floating-point roundoff through the loop's repeated `T = T @ new_T` composition. Traced this on a real scan-matching case (RTB's `PoseGraph.scanmatch`, killian dataset): deviation from orthonormality grows monotonically from ~1e-16 at iteration 1 to ~7.3e-15 after 15 iterations -- just over `ishom2()`'s `20*eps ≈ 4.44e-15` tolerance.
• When that check fails, downstream `SE2(T)` construction doesn't raise -- `arghandler()` returns `False` and execution falls through to the branch meant for `[x, y, theta]`-style list/tuple arguments, since a (3,3) array also satisfies `len(x) == 3`. This produces a confusing `TypeError: only 0-dimensional arrays can be converted to Python scalars` deep in `math.cos()`, nowhere near the real problem. (That fallthrough behavior is a separate, pre-existing `SE2.init` issue -- not addressed in this PR, flagging it here for visibility.)
• Fix: normalize the returned `T` with `trnorm2()` before returning, which is exactly the situation its own docstring describes ("prevent finite word length arithmetic causing transforms to become 'unnormalized'").

Test plan

• `pytest tests/base/test_transforms2d.py` -- 20 passed
• New `test_icp2d_returns_valid_se2` (seed fixed for determinism) confirmed to fail against the pre-fix implementation and pass with the fix
• Verified `RTB`'s `PoseGraph.scanmatch(100, 101)` on the real killian.g2o lidar dataset, which surfaced this, now runs end-to-end

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!