Improve random state management

[tautomer]

This PR addresses random state management issues in #80 to improve reproducibility.

Implementation

• RNGManager class: Centralized wrapper managing NumPy and PyTorch generators with synchronized seeding
• Single seed per campaign: One seed initialization ensures consistent randomness throughout campaign lifecycle
• Explicit RNG passing: Components (Campaign, Optimizer, ExpDesigner, Simulator) accept RNG instances explicitly instead of global state resets
• State persistence: Campaign and Optimizer save/load complete RNG state (seed + generator states)
• Ownership tracking: _owns_rng flag distinguishes owned vs shared RNG instances
• Temporary seed overrides: Context manager allows manual_seed parameter without global pollution
• Backward compatibility: obsidian.USE_OLD_RNG_CONTROL toggle for legacy behavior
• Comprehensive testing: 11 tests covering RNG lifecycle, state persistence, and reproducibility

Other changes

Some other improvements were made while this new RNG control was implemented.

1. Introduced direct optimizer options passthrough to model fitting routine in campaign.fit. You can now do

   campaign.fit({"max_attempts": 10, "multi_starts": True})

   to fit with model with 10 different restarts.

2. Introduced direct optimizer options passthrough to acquisition function optimization in optimizer.suggest

   campaign.optimizer.suggest(
       ...,
       optim_options={"method": "Powell", "maxiter": 100}
   )

   This will tell scipy.optimize.minimize to use Powell and maximum iteration of 100.

Current status

The following tests have been performed, both on single objective (modified from Simple single objective.ipynb)

1. The new code, with obsidian.USE_OLD_RNG_CONTROL=True produces the same results as the current main branch.
2. With new RNG control, two runs with the same initial random states achieve identical results with categorical parameters.
3. With new RNG control, two runs with the same initial random states achieve identical results without categorical parameters. (fixed in 1064d6f).