Defer SciPy and SymPy imports to cut `import spatialmath` time (~800 ms → ~150 ms)

[petercorke]

Follow-up to #198, which deferred matplotlib.pyplot until something actually plots. Its description noted that scipy.interpolate and sympy cost about as much. They are now the main remaining costs of import spatialmath.

Where the time goes

Measured with python -X importtime -c "import spatialmath" on Python 3.12, macOS, Apple M1 (8 GB), warm cache. Cold-start times will be higher.

Module	Cost	Imported by
scipy.interpolate + scipy.spatial.transform	~455 ms	base/quaternions.py:19, spline.py:14-15 (both at top level)
sympy (when installed)	~200 ms	base/symbolic.py:19, via base/argcheck.py
matplotlib (no pyplot)	~100 ms	geom2d.py:13 from matplotlib.path import Path. #198 kept this on purpose because Polygon2 needs it for its geometry
numpy	~40 ms
Total	~800 ms

SciPy: nothing needs it at import time

• base/quaternions.py: the only user is qrand(). When the angle is limited, it does inverse-CDF sampling through _generate_inv_cdf_sin_squared_interp(), which is already @lru_cache'd. Every user pays ~440 ms for this one option. qslerp and the rest of the module don't use SciPy. Fix: move the import into that function, and make the -> interpolate.interp1d annotation a string or put it under TYPE_CHECKING.
• spline.py: CubicSpline, RotationSpline, Rotation and BSpline are only used inside methods. Fix: import them inside those methods.
• Both must change. If only one does, the other still loads scipy.interpolate. SciPy stays a hard dependency; only when it loads changes.
• Fit b spline se3 #129 (open) adds more top-level SciPy imports to spline.py (scipy.optimize, splrep) and would need the same treatment.

SymPy: avoid importing it just to check types

SymPy is optional. base/symbolic.py imports it inside a try so that issymbol() can check for sympy.Expr. But if "sympy" not in sys.modules, no argument can possibly be a SymPy expression. So issymbol() can return False without importing SymPy, and only import it once the user has imported SymPy themselves. This is a separate change from the SciPy one because argcheck is on almost every code path, so it needs its own testing (including the :SymPy: supported paths).

Expected result

About 150 ms for import spatialmath, and none of SciPy or SymPy loaded until a feature actually uses them. A regression test along the lines of #198's could assert that neither scipy nor sympy is in sys.modules after import spatialmath.