Stabilize periodic GFN-FF optimization

[TyBalduf]

Potential fix for #1382. LLM generated and haven't gotten a chance to carefully look over it, but figured it might at least serve as the base for fixing this issue.

LLM short summary:

The periodic L-BFGS implementation is also corrected to apply wrapped
history in the required order, reject nonpositive-curvature updates, retain
the previous displacement correctly, enforce the configured maximum step,
and honor optimization level, maximum-cycle, and maximum-displacement input.
Focused regression tests cover history wrapping and curvature rejection.