backtest: implement walk-forward optimization and grid search

[marwinsteiner]

Summary

Implements src/sysls/backtest/optimize.py with walk-forward analysis, grid search, and parameter optimization for the vectorized backtester.

• ParameterGrid: Generates Cartesian product combinations from dict[str, list[Any]] via itertools.product. Iterable with __len__.
• GridSearchResult / WalkForwardSplit / WalkForwardResult: Frozen Pydantic v2 models for optimization results with JSON serialization support.
• grid_search(): Evaluates every parameter combination via run_vectorized_backtest, sorts by chosen metric (ascending for max_drawdown, descending otherwise). Validates metric against BacktestResult fields.
• TimeSeriesSplit: Expanding-window time-series cross-validation with configurable train_ratio and n_splits. Last split absorbs integer division remainder.
• walk_forward(): For each split, grid-searches training data then backtests OOS with best params. Concatenates and scales OOS equity curves for combined performance metrics.

Key design choices

• Pure computation, no async code (as required by CLAUDE.md for vectorized backtest path)
• BacktestResult imported at runtime (Pydantic needs it for field resolution)
• run_vectorized_backtest and summarize_backtest imported at runtime in function bodies
• structlog for logging, no print statements
• All type hints, Google-style docstrings on public APIs

Commit discipline (7 atomic commits)

1. Wireframe with stubs and type annotations
2. ParameterGrid with Cartesian product iteration
3. Pydantic models (GridSearchResult, WalkForwardSplit, WalkForwardResult)
4. grid_search() with metric-based optimization
5. TimeSeriesSplit expanding window helper
6. walk_forward() with expanding window optimization
7. Edge case tests and cleanup

Test plan

• 37 tests in tests/backtest/test_optimize.py - all passing
• Full suite: 688 passed, 10 skipped (Windows-only arcticdb skips)
• ruff check src/ tests/ - all passed
• ruff format --check src/ tests/ - all passed
• ParameterGrid: basic grid, single/empty/multi-param, length, re-iteration
• Pydantic models: construction, immutability, JSON round-trip
• grid_search: basic search, metric sorting, max_drawdown ascending, invalid metric error
• TimeSeriesSplit: expanding window, no overlap, coverage, invalid inputs
• walk_forward: basic, OOS concatenation, combined metrics, equity continuity, single split
• Edge cases: multi-param MA crossover, costs pass-through, 3-param grid, train_ratio validation