Adopt the modern-di integration kit

[lesnik512]

Summary

• Composes modern_di.integrations (shipped in modern-di 2.28.0) instead of hand-rolling the same logic — zero change to public API or test behavior.
• build_di_container's isinstance-over-tuple connection dispatch is replaced by integrations.classify_connection; the manual try/finally around the child container is replaced by Container's own async with protocol.
• Dependency's field changes from a raw dependency to marker: integrations.Marker[T_co]; its __call__ becomes self.marker.resolve(request_container). FromDI constructs Dependency(integrations.Marker(dependency)) — its own public signature is unchanged.
• pyproject.toml's modern-di floor bumped to >=2.28.0,<3.
• architecture/container-lifecycle.md and architecture/dependency-resolution.md promoted to describe the new internals.

This is a native-DI adapter (FastAPI's own Depends drives resolution), so — unlike the just-shipped modern-di-starlette conversion (modern-di-starlette 2.2.0) — there's no @inject decorator to replace, no parse_markers/resolve_markers usage, and no dead imports result. Second of 13 planned adapter conversions.

Design background: planning/changes/2026-07-13.01-adopt-integration-kit.md.

Test plan

• just test-ci — 8 passed, 100% line coverage
• just lint-ci — ruff format/check, ty check, check-planning all clean
• No test file touched anywhere in this branch — the existing suite is the regression check, and it passed unmodified after every task
• Built via subagent-driven-development: 5 planned tasks, each with an independent spec+quality review; a whole-branch review before merge — verdict "Ready to merge: Yes", zero findings at any severity