feat: retry + timeout + uniform errors for LLM calls (closes #28)

[PRATHAMESH75]

Summary

Closes #28. Adds resilient outbound handling for LLM provider calls — retry with backoff, per-request timeout, and a single actionable error surface — for every LLMClient-based agent (report, ontology, tools).

Stacked on #32 (feat/20) → #31 → #30. Merge order: #30 → #31 → #32 → #28.
Note: this is the outbound provider direction (429/timeout to the LLM). Incoming HTTP rate limiting is already wired via flask-limiter on the routes.

Changes

• _request_with_retries — retries transient provider errors (RateLimitError, APITimeoutError, APIConnectionError, InternalServerError) with exponential backoff; non-transient errors (bad request, auth) raise immediately (no pointless retries).
• Per-request timeout passed to the OpenAI client (primary + fallback).
• Order of resilience: primary model (with retries) → backup model (with retries, from Per-task model/provider config + fallback #15) → give up.
• LLMError — one user-facing, actionable failure surface raised when everything is exhausted. Subclasses RuntimeError and embeds the original detail + model name, so existing except Exception catchers and the Per-task model/provider config + fallback #15 fallback tests keep working unchanged.
• Config: LLM_MAX_RETRIES, LLM_RETRY_INITIAL_DELAY, LLM_RETRY_MAX_DELAY, LLM_TIMEOUT (documented in .env.example).

Acceptance criteria

• Provider 429/timeout handled gracefully across all (LLMClient-based) agents
• User-facing errors are clear and actionable (LLMError with retry/provider guidance)

Testing

• New tests/test_llm_retry.py (5 tests): retry-then-succeed, exhausted→LLMError with actionable message, non-transient not retried, transient-on-primary→fallback succeeds, timeout class membership.
• Full backend suite: 54 passed.

Scope note

The two raw-OpenAI generators (oasis_profile_generator, simulation_config_generator) keep their own existing retry loops; this PR centralizes resilience for the shared LLMClient path. Folding those onto LLMClient is a reasonable follow-up.

🤖 Generated with Claude Code