Enhancement: SE3 constructor should accept an (N, 4, 4) ndarray as N poses #236

Summary

SE3 can be built from a single 4x4 array, from a list of 4x4 arrays (its own internal representation for a sequence of poses), or from other SE3 objects, but not from a stacked (N, 4, 4) ndarray. That is the natural way to hold a pose sequence in NumPy-based code (for example the output of np.stack, or a trajectory loaded from a file), and it is the form that a number of Robotics Toolbox for Python (RTB) functions accept for pose trajectories.

Reproduction

import numpy as np
from spatialmath import SE3

A = SE3.Rx(0.3).A
arr = np.array([A, A, A])        # shape (3, 4, 4)

SE3(arr)
# ValueError: bad argument to constructor

s = SE3(arr, check=False)
len(s), s.A.shape
# (1, (3, 4, 4))   <- silently a malformed object: length 1, but .A is (3, 4, 4)

len(SE3([A, A, A]))              # 3, a list of matrices works

Tested with spatialmath-python 1.1.18.

Why it matters

• With the default check=True the error message does not say what is accepted.
• With check=False the constructor silently returns a malformed object. In RTB this turned a user's (N, 4, 4) trajectory into an object that failed deep inside an IK solver with ValueError: operands could not be broadcast together with shapes (3,4) (3,), far from the cause.

Suggested behaviour

1. Accept an ndarray with ndim == 3 and shape[1:] == (4, 4) as a sequence of N poses, equivalent to SE3(list(arr)).
2. For other unsupported shapes, raise an error that says which forms are accepted. Ideally this validation of the array shape (not the matrix contents) would also happen with check=False, which would skip only the expensive checks that the matrices are valid SE(3).

This is an enhancement request, not a regression. RTB works around it by converting the array to a list of matrices before calling the constructor, so there is no urgency.