feat: add Constraint dataclass and feasibility filtering

[jc-macdonald]

Summary

Adds constraint handling for feasibility filtering of designs (#74).

Changes

• Constraint dataclass in protocols.py: frozen dataclass with name, observable, op, threshold. Validates operator on construction (>=, <=, >, <, ==, !=). Includes a .check(value) method for scalar tests.
• ResultsTable.feasible(constraints): returns a boolean NDArray mask — True for rows where all constraints are satisfied. Supports both observable and annotation columns.
• feasibility_filter(constraints) in study.py: creates a Phase.filter_fn-compatible callable that keeps only feasible designs, enabling constraint-driven phase transitions.
• Exported Constraint and feasibility_filter from the public API.

Example usage

from trade_study import Constraint, Phase, feasibility_filter

constraints = [
    Constraint("min_conversion", "conversion", ">=", 0.5),
    Constraint("budget", "cost", "<=", 100.0),
]

Phase(
    name="screen",
    grid=grid,
    filter_fn=feasibility_filter(constraints),
)

Tests

• 17 new tests covering Constraint creation, operator validation, all 6 comparison operators, ResultsTable.feasible with single/multiple/annotation constraints, and feasibility_filter standalone and inside a Study pipeline
• 258 total tests passing, 99.57% coverage

Closes #74