fix(ioc): drop type-lie in eager-DI provider construction

[lesnik512]

Summary

modern-di's Factory.resolve() eagerly resolves all provider_kwargs. The 0.3.x current_provider factory declared both gitlab_provider and github_provider as kwargs, so both got constructed on every resolution. When provider=github, _build_gitlab_provider(settings, gitlab_client) was called with settings.project_id=None → GitLabProvider(project_id=None) — a field typed int holding None, suppressed with # ty: ignore. Harmless today (no method touches self.project_id at construction), but trivially broken if anyone adds a __post_init__ validator or eager field access.

This PR collapses the per-provider factories into a single _build_current_provider that takes both clients and constructs only the active provider, with asserts that narrow the Optional fields and document the Settings._resolve_provider validator invariant.

• Delete _build_gitlab_provider, _build_github_provider, _close_gitlab_provider, _close_github_provider
• Delete ProvidersGroup.gitlab_provider, ProvidersGroup.github_provider
• Consolidate finalizers into one _close_client attached to each client factory's cache_settings (lifecycle target moved from provider to client)
• _build_current_provider now takes (settings, gitlab_client, github_client); the unused client still resolves (eager Factory) but httpx2 connection pools are lazy

420 tests pass, 100% coverage, lint clean — existing test_ioc.py tests resolve current_provider and check the type, which still works under the new shape.

Test plan

• just test — 420 pass (same total, no test removed/added)
• just lint-ci — clean
• Existing test_container_resolves_github_provider_when_settings_provider_is_github + ..._gitlab_..._is_gitlab confirm both branches dispatch correctly
• Merge — bugfix/ prefix triggers dogfood auto-tag (0.3.2 → 0.3.3)
• Manually cut 0.3.3 GitHub release to publish to PyPI (optional — internal-only refactor, no user-facing change)

🤖 Generated with Claude Code