feat: fit Ebbinghaus and exponential forgetting curves to recall@T data

[Priyanshu-byte-coder]

Summary

Implements forgetting-curve fitting requested in issue #6.

New function: evaluation/stats.py — fit_forgetting_curve()

Accepts (checkpoints, recalls) and fits two models:

Exponential decay R(t) = a · exp(−k · t)

• Fitted via log-linear least squares (no extra dependencies)
• Returns: a, k, half_life = ln(2) / k, R²

Ebbinghaus forgetting curve R(t) = exp(−t̂ / (S · √(1 + t̂)))

• t̂ = t / t_max (normalised)
• S = stability constant; higher = slower forgetting
• Fitted via fine grid search over S ∈ [0.01, 20] (avoids scipy dependency)
• Returns: stability, half_life (numerical), R²

CLI flag: --fit-curves

python main.py --fit-curves
python main.py --seeds 5 --fit-curves

Example output:

FORGETTING CURVE FIT  (Ebbinghaus + Exponential)
-----------------------------------------------------------------
  naive
    Exponential  k=0.004200  half-life=165.0 turns  R²=0.921
    Ebbinghaus   S=2.1000    half-life=142.3 turns  R²=0.944
  rag
    Exponential  k=0.006800  half-life=101.9 turns  R²=0.887
    Ebbinghaus   S=1.3500    half-life=88.5 turns   R²=0.912
  cascading
    Exponential  k=0.003100  half-life=223.6 turns  R²=0.956
    Ebbinghaus   S=2.8200    half-life=195.0 turns  R²=0.971

Half-life and stability give researchers a single interpretable scalar per backend rather than requiring point-by-point curve inspection.

Closes #6

[Neal006]

Manually squash-merged into main (8a58920) after resolving conflict with PR #18's --scenario argument. Both --fit-curves and --scenario flags are now live on main. Thanks @Priyanshu-byte-coder!