Harden loss functions and raise evaluation rigor

[v-atharva]

Summary

• Correctness (loss_functions.py): fix + vs * typo in custom_loss_13/14/15/16 (entropy and HHI were added to their lambdas instead of scaled by them); clamp 1 + r to >= eps in log_return / log_sharpe / log_sortino so returns near -1 don't produce non-positive log args; validate temp/beta > 0 at entry of every softplus-based loss and eps-floor the beta = 1/temp in Rockafellar CVaR.
• Tests: new tests/unit/training/test_loss_functions.py (99 cases) — registry-driven shape/finiteness/gradient coverage for every registered loss, plus targeted regressions for each fix and semantic checks (HHI uniform=0 / concentrated=1, CVaR top-k matches empirical tail mean, log-return monotonicity).
• Evaluation: add MinVarianceCalculator (long-only QP via cvxopt, optional linear shrinkage) and EqualRiskContribCalculator (Maillard-Roncalli-Teiletche fixed point); add stationary-block bootstrap helpers bootstrap_metric_ci and bootstrap_paired_diff_ci, exposed via Evaluator.calc_metric_performance_ci and calc_paired_diff_ci; register turnover in MetricLibrary and add Evaluator.calc_turnover_for_all and calc_cost_adjusted_daily_rets.
• Training robustness: Trainer now supports periodic checkpointing (controlled by train_hparams.checkpoint_every and checkpoint_keep_last) persisting model, optimizer, scheduler, RNG, and early-stopping state atomically; new Trainer.resume_from() and find_latest_checkpoint(); scripts/run_multi_train.py accepts --resume and writes per-seed checkpoints under artifacts/results/checkpoints/<combo-seed>/.

Test plan

• pytest tests/unit/training/test_loss_functions.py — 99 / 99 pass
• Full pytest tests/unit — 306 pass; 5 pre-existing failures on main (unrelated: test_loading.py, test_io.py)
• Smoke-tested MinVarianceCalculator + EqualRiskContribCalculator on synthetic returns — weights sum to 1, long-only, ERC risk contributions equal to ~1e-6
• Smoke-tested bootstrap CI + paired diff + calc_metric_performance_ci on synthetic data
• Smoke-tested Evaluator.calc_turnover_for_all and calc_cost_adjusted_daily_rets — turnover in [0, 1], cost-adjusted mean < gross mean
• Smoke-tested Trainer checkpoint/resume: 4-epoch run → kill → resume from epoch 4 → 2 more epochs; loss continues decreasing across the boundary, oldest checkpoint pruned per keep_last
• Re-run a real scripts.run_multi_train grid once dependencies (statsmodels etc.) are installed on the target env