More cleanup of `general_` functions

This extracts as much of the common logic from the nd-transforms as possible into generic functions. With the transform-specific logic handled by the exec functor. I've also tried to abstract all of the data movement and vector-type reordering into overloads of the various copy_ functions.

This does currently keep r2c and c2r as their own special cases since they are subtly different in a few ways.

test.py

@@ -58,9 +58,18 @@ def irfft_scipy(a, axis, inorm=0, out=None, nthreads=1):
 tol = {np.float32: 6e-7, np.float64: 1.5e-15, np.longfloat: 1e-18}
 ctype = {np.float32: np.complex64, np.float64: np.complex128, np.longfloat: np.longcomplex}
 
+import platform
+on_wsl = "microsoft" in platform.uname()[3].lower()
+true_long_double = (np.longfloat != np.float64 and not on_wsl)
+dtypes = [np.float32, np.float64,
+          pytest.param(np.longfloat, marks=pytest.mark.xfail(
+              not true_long_double,
+              reason="Long double doesn't offer more precision"))]
+
+
 @pmp("len", len1D)
 @pmp("inorm", [0, 1, 2])
-@pmp("dtype", [np.float32, np.float64, np.longfloat])
+@pmp("dtype", dtypes)
 def test1D(len, inorm, dtype):
     a = np.random.rand(len)-0.5 + 1j*np.random.rand(len)-0.5j
     a = a.astype(ctype[dtype])
@@ -199,7 +208,7 @@ def test_genuine_hartley_2D(shp, axes):
 @pmp("len", len1D)
 @pmp("inorm", [0, 1])  # inorm==2 not needed, tested via inverse
 @pmp("type", [1, 2, 3, 4])
-@pmp("dtype", [np.float32, np.float64, np.longfloat])
+@pmp("dtype", dtypes)
 def testdcst1D(len, inorm, type, dtype):
     a = (np.random.rand(len)-0.5).astype(dtype)
     eps = tol[dtype]