Multi-D/Grain Sample Triangles

[kennethshsu]

I do think an improvement is that if we can get various style of dataset ready it could be good. For example, simple triangle, multi-D triangle, with combinations of (M,Q,Y)x(M,Q,Y) combinations all readily available.

Originally posted by @kennethshsu in #1205 (comment)

Coming out of this comment, we should create sample triangles that's 2D, 3D, 4D, and have varying (MxQxY) x (MxQxY) for better testing. We are always creating synetic sample triangles manually for various examples.

[ppcvote]

This is the follow-up I offered to take in #1205, so consider the word said — I would like to pick it up, with the same caveat as there: you and the other owners sight the expected values.

Before writing anything, three scoping questions so the fixtures earn their keep rather than becoming another thing to maintain:

1. Where should they live? Options I see: (a) data files under chainladder/utils/data/ registered like the existing samples, loadable via load_sample, or (b) a factory in a test helper module that builds them on demand (make_triangle(index_dims=2, origin_grain="Q", development_grain="M")). I lean factory: the grain matrix is 16+ combinations before index dimensionality enters, and generated-on-demand keeps the wheel small and the expected values derivable rather than frozen. But if you want them documentable and usable in examples like raa/clrd, files win. Which audience matters more?

2. Should S be in the matrix from day one? fix(parallelogram): support quarterly and semiannual origin grains #1205 made semiannual origins work for ParallelogramOLF, but S support elsewhere is uneven, so an (S)x(...) fixture would immediately surface failures outside the parallelogram path. That might be exactly the point — or noise you would rather stage. Your call.

3. Is ParallelogramOLF daily grain ValueError after leap year #1219 the first consumer? The daily-approximation leap bug needs a triangle whose origin starts right after a leap year to reproduce, which is precisely the kind of awkward shape nobody hand-builds twice. If the fixture design covers "origin starts at an arbitrary date" it pays for itself on that issue alone.

If the answers are "factory, yes, yes" I can have a first PR up with the factory, the (M,Q,S,Y)x(M,Q,S,Y) matrix behind it, and #1205's parallelogram tests rewritten on top of it as the proof of use.