fix: normalize runtime lm objects and retries

[magix022]

Rationale

Fractal resolves provider configuration into the language models passed to PredictRLM. Those runtime LMs need to be concrete dspy.LM instances so credentials, model labels, retry behavior, and downstream DSPy/PredictRLM behavior are consistent across configured providers, explicit --lm values, and interactive model/provider changes.

This change also adds a configurable LM retry default so transient provider failures can be retried consistently without leaking retry settings into RLM inputs or requiring users to pass lower-level DSPy options manually.

Summary

• Normalize runtime model strings and provider selections into concrete dspy.LM objects.
• Build API-backed provider LMs with resolved API keys instead of relying on provider env lookup later in the call path.
• Add defaults.num_retries and FRACTAL_LM_NUM_RETRIES support.
• Preserve configured retry counts on startup, explicit --lm, configured sub-models, and interactive /model or /provider swaps.
• Update config documentation and focused tests for LM construction, retry propagation, and runtime labels.

Test Plan

• Run uv run pytest.
• Confirm full suite passes: 283 passed, 1 deselected.
• Confirm defaults.num_retries reaches the underlying DSPy completion retry budget.
• Confirm retry settings are not passed as PredictRLM input fields.

[codecov]

Codecov Report

❌ Patch coverage is 81.94444% with 13 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/fractal/config.py	66.66%	6 Missing ⚠️
src/fractal/runtime.py	78.94%	4 Missing ⚠️
src/fractal/runtime_lms.py	91.30%	2 Missing ⚠️
src/fractal/providers.py	91.66%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!