Cleanup: remove runtime-init globals, add public reset_logging() (follow-up to #755)

[nabinchha]

Summary

Follow-up cleanup to #755 (which fixes #388). #755 is correct and should ship as-is, but it leaves two module-global boolean latches in the runtime-init path that force tests to monkeypatch private globals:

• data_designer.logging._logging_configured
• data_designer.interface.data_designer._interface_runtime_initialized

This issue tracks removing both in favor of observable/idempotent state, and adding a public reset_logging() API.

Goals

1. Derive "did Data Designer configure logging?" from observable handler state instead of a private boolean latch.
2. Remove _interface_runtime_initialized entirely — the work it guards (configure_logging() and resolve_seed_default_model_settings()) is already idempotent (the CLI's ensure_cli_default_model_settings() already depends on this).
3. Add a public reset_logging() so external apps can undo Data Designer's logging setup — and so tests need zero private patching.
4. As a side effect, fix the auto_configure_logging "first-construction-wins" footgun: with the memo gone, the flag takes effect on every DataDesigner() instead of only the first one in a process.

Design sketch

logging.py

• Mark DD-installed handlers via type (_DataDesignerStreamHandler / _DataDesignerFileHandler sharing a _DataDesignerManagedHandler mixin); _create_handler() returns these.
• is_logging_configured() → any(isinstance(h, _DataDesignerManagedHandler) for h in logging.getLogger().handlers). Delete _logging_configured.
• Add public reset_logging() that surgically removes only DD-managed handlers (leaving foreign handlers such as pytest's caplog capture handler intact) and resets root/data_designer levels to defaults.

data_designer.py

• Simplify _initialize_interface_runtime() to be idempotent and drop _interface_runtime_initialized + the early-return guard.
• Trim the "first construction only" caveat from the auto_configure_logging docstring.

Behavior changes to note in the follow-up PR

• is_logging_configured() shifts from "was ever configured this process" to "is currently configured" (only observable if someone configures then manually clears handlers; the sole internal consumer checks once at construction).
• Runtime-init side effects are re-evaluated per construction (idempotent; no new log spam).
• auto_configure_logging now applies per-instance.

Test plan (zero private patching)

• Shared autouse fixture calling reset_logging() before and after each test (replaces all monkeypatch.setattr(..., "_logging_configured"/"_interface_runtime_initialized", ...) and try/finally blocks from fix: preserve preconfigured logging #755). caplog stays intact because the reset is surgical.
• reset_logging() contract tests: clears DD handlers / flips is_logging_configured() to False, preserves foreign handlers, idempotent no-op on a clean process.
• Interface tests: delete test_initialize_interface_runtime_runs_once (property no longer holds) → replace with an idempotency test; add a per-instance auto_configure_logging test; optionally a subprocess fresh-interpreter configure_logging() called before DataDesigner() is silently overwritten #388 regression test (reuse _run_python_snippet from test_lazy_imports.py).

Docs

• Document reset_logging() as the public "undo" affordance that pairs with configure_logging().

Notes

• Full design + decisions log: plans/logging-runtime-init-cleanup.md.
• Cost: ~3 extra Path.exists() stats per DataDesigner() construction — negligible (heavyweight top-level object, never hot-looped).
• Do not start until fix: preserve preconfigured logging #755 is merged.

Follow-up to #755. Closes #388's remaining tech debt.