Prepare ml4t-models 0.1.0 stable release

[stefan-jansen]

What changed

• freezes the supported public API and validates estimator, configuration, dtype, lifecycle, and persistence contracts
• corrects financial objectives, portfolio normalization, neural refit behavior, checkpoint persistence, and deterministic CUDA execution
• adds release-candidate construction, reproducibility, installed-wheel, coverage, dependency, hardware, and performance qualification gates
• updates packaging metadata, stable-version identity, security policy, installation guidance, and release documentation

Why

The stable-readiness audit found 44 issues, including 39 release blockers. This change closes the code and local-evidence findings and moves the remaining operating-system, Python, CUDA, performance, release, and deployment evidence into enforceable GitHub workflows.

Impact

This prepares ml4t-models 0.1.0 for a stable release. It tightens validation and persistence behavior, makes public errors actionable, and requires one immutable candidate artifact to pass the release gates before publication.

Validation

• 398 source tests passed
• 378 tests passed against the installed candidate wheel
• line coverage: 95.49%; branch coverage: 85.20%
• Ruff, Ruff format, ty, actionlint, strict MkDocs, current and minimum dependency audits passed
• reproducible wheel comparison and package metadata checks passed
• RTX 3090 CUDA qualification, deterministic replay, canonical performance, and isolated scaling checks passed locally

The PR remains draft until the supported GitHub-hosted matrix and protected remote release workflow have passed.