fix: enable LLM retry without fallbacks and surface SDK retries

[ngoclam9415]

Summary

Fixes a bug where an Azure APIConnectionError exited the STAR loop after ~7.5 minutes with no visible retry attempts, despite the LLMCaller having retry logic. Three converging bugs were responsible:

1. LLMCaller retry was opt-in via fallback_providers — primary provider had no retry path. Now retry/backoff always runs; failover stays gated on fallback_providers.
2. _is_transient_error did not classify "Connection error." as transient (no matching keyword) and could not see openai.APIConnectionError buried in __cause__. Added "connection" to _TRANSIENT_KEYWORDS and a __cause__-chain walk that detects OpenAI SDK exception class names.
3. STAR loop exited on any exception. Added one bounded retry per iteration for transient LLM errors (sync + async paths), with full tracebacks via exc_info=True on all error sites.

Observability improvements

• New make_logging_http_client wired into OpenAI/Azure/Moonshot providers — emits LLM HTTP request per attempt, surfacing SDK-internal retries that were previously silent (those took ~7.5 min in the original incident).
• New structured events: llm_retries_exhausted, STAR transient retry warnings, OpenAI-compatible API connection error with traceback.
• All exception logs gained error_type + exc_info=True.

Files changed

• dana/core/llm/llm_caller.py — always-retry path, expanded transient detection, public is_transient_llm_error
• dana/common/llm/providers/openai_compatible_base.py — APIConnectionError handling, logging http client helper
• dana/common/llm/providers/{azure,openai,moonshot}.py — wire logging http client
• dana/core/agent/base_star_agent.py — STAR transient retry (sync+async), tracebacks
• dana/core/agent/star_agent_streaming.py — traceback + transient flag in error log
• tests/unit/core/test_llm_caller_failover.py — regression tests for connection-error classification, __cause__ walk, no-fallback retry

Test plan

• pytest tests/unit/core/test_llm_caller_failover.py — 24 passed
• pytest tests/unit — 1055 passed, 20 skipped
• pytest tests/regression — 44 passed, 18 skipped
• tests/live/llm/test_multi_modal.py against real API keys — all providers (anthropic, openai, azure, gemini, gemini_openai, anthropic_like, moonshot) returned OK on every supported modality
• Verify in a real Azure flake scenario that retry logs appear and the STAR loop survives