fix: RITS/hosted_vllm pricing tolerance, AppWorld init lock, env-conf

[korenLazar]

…igurable timeouts, child-process env forwarding

Small set of focused fixes uncovered while running CE-Manager + Mnemex sweeps on AppWorld with RITS-hosted models. Each fix is independently useful for any user running:

• benchmarks with parallel sessions on AppWorld,
• models that LiteLLM has no pricing for (e.g. hosted_vllm/, rits/),
• subprocess (venv) runners with slow model-health endpoints,
• subprocess runners that need PYTHONPATH or locale forwarded.

Bugfixes:

• benchmarks/appworld/appworld_eval.py: serialise AppWorld(init) under a class-level threading.Lock. AppWorld's init clears a global cache dict that races under parallel workers, producing 'dictionary changed size during iteration'. New tests/benchmarks/test_appworld_session_init_lock.py.
• integrations/litellm/trace_logger.py: guard against ctx is None or ctx.otel_context is None in start_span_context_from_kwargs. Avoids AttributeError when LiteLLM emits trace events outside an Exgentic context. Coverage in tests/observers/test_otel.py.
• observers/handlers/results.py + utils/cost.py: tolerate models that litellm.cost_calculator has no pricing for (currently hosted_vllm/* and rits/*). Previously crashed the entire results pipeline with 'No pricing info found for model'; now returns zero-cost and logs a warning. Coverage in tests/utils/test_cost.py.

Configurability:

• adapters/runners/venv.py: EXGENTIC_VENV_HEALTH_TIMEOUT and EXGENTIC_VENV_TRANSPORT_TIMEOUT override the previously-hardcoded defaults (360s / 600s), needed for slow inference endpoints.
• agents/smolagents/base_instance.py: EXGENTIC_MODEL_HEALTH_TIMEOUT flows into check_model_accessible_sync.
• adapters/runners/utils.py: forward PYTHONPATH, HOME, USER, LANG, LC_ALL, and CE_MANAGER* env vars to subprocess runners so optional packages and locale survive the venv hop. Coverage in tests/adapters/runners/test_utils.py.