feat: add preflight checks, --dry-run flag, and credential timeout resilience

[placerda]

Summary

Generalizable improvements to agentops eval run based on a real-world RAG evaluation testing session where the workflow failed multiple times on issues that could be detected upfront. Addresses the friction points without adding project-specific code.

Changes

Phase 1 — Pre-flight check system

New src/agentops/services/preflight.py module with 4 checks that run before backend execution in runner.py:

1. SDK imports — azure-identity and azure-ai-evaluation. Fails with pip install hints.
2. Environment variables — AZURE_OPENAI_ENDPOINT/AZURE_OPENAI_DEPLOYMENT for AI-assisted evaluators; AZURE_AI_FOUNDRY_PROJECT_ENDPOINT for safety evaluators.
3. Credential warm-up — calls DefaultAzureCredential.get_token() once so subsequent evaluator calls use the MSAL cache instead of each cold-starting az.cmd.
4. Endpoint reachability — lightweight HTTP HEAD (10s timeout) on remote-mode endpoints.

All checks run together and collect errors — failures are reported as a single numbered list so users can fix everything in one pass.

Phase 2 — --dry-run / -n flag on eval run

Runs pre-flight checks only. Exit 0 if all pass, 1 if any fail. Useful for CI gating and fast feedback loops.

Phase 3 — Credential timeout resilience

Added process_timeout=30 to all 3 DefaultAzureCredential sites:

• eval_engine.py::_default_credential()
• foundry_backend.py::_acquire_token()
• foundry_backend.py::_invoke_model_direct()

The default 10s is insufficient for Windows az.cmd cold starts and was producing intermittent AzureCliCredential: Failed to invoke the Azure CLI errors.

Phase 4 — Tests

New tests/unit/test_preflight.py with 14 tests covering:

• PreflightReport formatting
• Evaluator classification helpers
• Local-only bundles skip Azure checks
• Missing env vars reported
• Missing SDK reported with pip install hint
• Endpoint unreachable detected
• Multiple errors collected at once
• HTTP endpoint success path

Test Results

253 passed, 1 skipped in 11.93s

All existing 239 tests continue to pass; 14 new preflight tests added.

[placerda]

Cancelling for now, we'll prioritize #107