[barrosyan]

Addresses items 2 and 3 of #27. Item 1 is left alone on purpose (see below), so I am not using Fixes.

Changes

• G = E / (2 * (1 - nu)) is now E / (2 * (1 + nu)). At nu = 0.33 the old value was about 1.985x too high.
• The torsion constant J was I1 + I2 (the polar moment of area), which is only the torsion constant of a circular section. It now uses the St. Venant constant of the rectangle, J = (a c^3 / 3) (1 - (192 / pi^5) (c / a) sum_{n odd} tanh(n pi a / (2 c)) / n^5) with a the long side and c the short side (Timoshenko and Goodier). It is 1.19x smaller for a square and 3.17x smaller for a 1:3 section. The series uses 100 odd terms (error below 1e-10).
• Docstring note on both. ae, ei1, ei2, mass_per_length and tmmi_per_length are unchanged (the polar moment is correct for the rotary inertia term).

Tests

New tests/beam_test.py (11 tests):

• J against the classical torsion coefficients for side ratios 1, 1.5, 2, 3 and 10 (0.1406, 0.1958, 0.2287, 0.2633, 0.3123), and symmetry in b, h;
• G through jg: the ratio of the values at nu = 0 and nu = 0.5 must be 3/2 (the old formula gave 1/2), and the value for a square section;
• the first torsion frequency of a 1 m, 10 x 30 mm aluminum cantilever (40 elements) against sqrt(G J / (rho Ip)) / (4 L): within 0.5% (the issue's reproducer gives 438.7 Hz by the closed form and 1099.7 Hz with the old code);
• the other returned properties are unchanged.

With the old code 8 of the 11 fail; with the change all pass, and the whole suite passes (88 tests, including the 77 existing ones). Models that use rect_beam_props (System.beam, beamkm_2d, the frame_wing demo) will see lower torsional stiffness, as physics requires. The 2D beam keeps only bending and axial dofs, so it is not affected.

About item 1 (bending stiffnesses)

I did not change I1 and I2. beamkm documents ei1 as the stiffness about bending axis 1, and with that reading I1 = b h^3 / 12 is consistent. The in-repo callers pass (width, height) as (b, h) and depend on it. If the intended reading is deflection along bend_direction_1, the cleaner fix is probably in the rect_beam_props docstring, but that is a decision for the maintainers (more detail in my comment on #27).