Implement a 1D adaptive sampling learner backed by BTreeMap/BTreeSet
internals, matching the semantics of Python's adaptive.Learner1D.

- Core Learner1D with tell/tell_pending/tell_many/ask/run methods
- LossManager priority queue with O(1) interval lookup + sorted access
- Built-in loss functions: Default, Uniform, Resolution, Curvature,
  Triangle, AbsMinLog, and Python callback support
- 44 comprehensive Python tests covering all functionality
- Python benchmark script for performance measurement

- Replace .expect() with error handling in python_callback_loss (CRITICAL)
- Store out-of-bounds data separately to prevent neighbor query pollution (HIGH)
- Keep x_scale constant as bounds span, not derived from data (HIGH)
- Expose loss(real) parameter in PyO3 binding (MEDIUM)
- Fix empty learner linspace to include right endpoint (MEDIUM)
- Add interval-based tiebreaking in ask() loss comparison (MEDIUM)
- Use iterator .any() for has_missing_bounds instead of Vec allocation (LOW)
- Avoid cloning in rebuild_scale via free function (LOW)
- Add 12 new tests covering OOB tells, callback exceptions, loss(real=False)