feat: add reference-free FluencyLLMEval and CoherenceLLMEval descriptors

[mostaphaelansari]

Summary

Closes #1801

Adds two new reference-free LLM-as-a-judge descriptors that evaluate text quality without requiring a reference dataset or ground-truth answer.

Descriptor	Classification	Reference needed?
FluencyLLMEval	FLUENT / NOT_FLUENT	❌ No
CoherenceLLMEval	COHERENT / INCOHERENT	❌ No

Motivation

Many real-world LLM monitoring setups don't have a golden reference to compare against. This PR fills that gap for two common quality dimensions — language fluency and logical coherence — following the same LLM-judge approach already used for ToxicityLLMEval, BiasLLMEval, PIILLMEval, and DeclineLLMEval (all of which are reference-free).

Changes

src/evidently/legacy/descriptors/llm_judges.py

• Added FluencyLLMEval: binary classification prompt (FLUENT / NOT_FLUENT) that checks grammar, natural phrasing, and readability
• Added CoherenceLLMEval: binary classification prompt (COHERENT / INCOHERENT) that checks logical organization and consistency

src/evidently/descriptors/generated_descriptors.py

• Added public factory functions for both descriptors with the same optional parameters as all existing evals (include_score, include_reasoning, uncertainty, alias, tests)

src/evidently/descriptors/__init__.py

• Exported FluencyLLMEval and CoherenceLLMEval

tests/features/test_llm_judge.py

• test_fluency_llm_eval: runs full pipeline via MockLLMWrapper (no API key required)
• test_coherence_llm_eval: runs full pipeline via MockLLMWrapper (no API key required)
• test_reference_free_evals_importable: verifies public import path from evidently.descriptors

Tests

All existing tests continue to pass, and 3 new tests were added:

10 passed, 6 warnings in 0.09s

Usage

from evidently.descriptors import FluencyLLMEval, CoherenceLLMEval

dataset.add_descriptors([
    FluencyLLMEval("response", provider="openai", model="gpt-4o-mini"),
    CoherenceLLMEval("response", provider="openai", model="gpt-4o-mini"),
])

Both descriptors also support include_score=True, include_reasoning=True, and uncertainty options.