Add bladeStiffenedShell constitutive model

[A-CGray]

This PR adds a new TACSBladeStiffenedShellConstitutive class, a new implementation of the old bladeFSDT class. It also adds a panel length constraint class to the pyTACS interface.

Constitutive model

Compared to the old bladeFSDT class, the new model contains a few improvements:

1. The panel and stiffener laminates can now be made up of an arbitrary number of ply angles as opposed to being hardcoded to [0, -45, 45, 90] deg laminates. This means you can model more complex laminates and makes it simpler to model a shell made from an isotropic material as you no longer need to supply 4 copies of the same isotropic ply when creating the constitutive object.
2. The buckling failure envelope is modified in a similar manner to Modify Tsai-Wu to work with safety factors #203 so that it can be used to ensure a specified buckling margin
3. I fixed a bug (I think) in the critical buckling loads calculation where the $D_{12}$ entry of the bending stiffness was being used instead of $D_{22}$ to compute the $D_2$ value in Stroud and Agranoff's equations.
4. Almost all of the calculations in the model have been rewritten in a way that computes contributions from the panel and stiffener as if they are distinct shell and beam constitutive models. My hope is that in the future, with a little more work, we could modify this model to create a stiffened shell model using any combination of existing TACS shell and beam constitutive models.

For example, this is how the stiffness matrix for the stiffened shell is computed:

// Compute the stiffness matrix
void TACSBladeStiffenedShellConstitutive::computeStiffness(TacsScalar C[]) {
  TacsScalar* A = &C[0];
  TacsScalar* B = &C[6];
  TacsScalar* D = &C[12];
  TacsScalar* As = &C[18];
  TacsScalar* drill = &C[21];

  // --- Zero out the C matrix ---
  memset(C, 0, this->NUM_TANGENT_STIFFNESS_ENTRIES * sizeof(TacsScalar));

  // Add the panel's stiffness contributions
  this->computePanelStiffness(C);

  // Compute the stiffener's beam stiffness matrix ,then transform it to a shell
  // stiffness matrix and add it
  TacsScalar Cstiff[TACSBeamConstitutive::NUM_TANGENT_STIFFNESS_ENTRIES];
  this->computeStiffenerStiffness(Cstiff);
  this->addStiffenerStiffness(Cstiff, C);
}

This code does not make any assumptions about the form of the panel or stiffener stiffness matrices and could easily be modified so that computePanelStiffness is actually a call to another shell constitutive model, and computeStiffenerStiffness is a call to a beam constitutive model.

Panel length constraints

I also implemented a panel length constraint class that is necessary to ensure the panel length values used in the buckling failure calculations are consistent with the true geometry of each TACS component. I used numba to speed up the panel length computation and its derivatives which means we would need to add Numba as a dependency of TACS. As far as I can tell this shouldn't be too much of a stability issue, numba has a pretty wide user-base, and is officially supported by organisations like Anaconda, Intel, and NVIDIA.

ToDo before merging:

• Merge Modify Tsai-Wu to work with safety factors #203
• Merge Unit testing tolerance improvements + Von Mises sensitivity fix #216
• Add numba dependency
• Fix Tsai-Wu derivative errors
• Add integration test for panel-length constraints
• Add integration test for BladeStiffenedShell
• Add BladeStiffenedShell example case?

[A-CGray]

@timryanb on further inspection, there was no issue with the Tsai-Wu sensitivities but there was an issue with the Von Mises sensitivities (not related to any changes in this PR) that I fixed. There were some issues with the finite difference tests for the blade stiffened model where the relative tolerance cannot be met because the absolute values of the derivatives are very small. To get around this I implemented a new type of pass/fail criteria for the unit tests based on numpy's assert_allclose. It basically acts like a relative tolerance when the values being compared are large but more like an absolute tolerance when the compared values are very small which fixes the issue I was facing.

This does change the criteria used for all the constitutive model unit tests, which I realise is a big change, and not related to the original purpose of this PR, so happy to make changes and possibly move this to another PR.

[A-CGray]

I stand corrected, there were a couple of uninitialized values causing problems only in the real tests when compiled in debug mode. I've fixed those and with the new assert_allclose check I was able to get all of the constitutive model tests passing with tighter tolerances

[timryanb]

Good find @A-CGray

The PR seems to drifting in scope a bit. Can we break the general test improvement/fixes into a separate dedicated PR?

[A-CGray]

@timryanb I think once #216 is merged, the tests here should pass and this PR will be ready to go

[timryanb]

@A-CGray I had to rename the benchmark file to get testflo -b to work since we already had a file called benchmark_analysis.py and testflo didn't like that.