feat: add LiteLLM backend - #1332

Summary

Adds LiteLLM as a first-class CLTK generative backend for synchronous and asynchronous NLP pipelines.

The backend uses LiteLLM's OpenAI-compatible Chat Completions endpoint, so a CLTK pipeline can select any model alias configured by the gateway, including aliases backed by OpenAI, Anthropic, Azure, or gateway fallback routes.

Changes

• add litellm to CLTK's supported backend configuration
• add LiteLLMBackendConfig with model alias, base URL, API key, temperature, and retry settings
• add synchronous and asynchronous LiteLLM connections
• route dependency parsing, morphosyntax, enrichment, and translation through LiteLLM
• preserve CLTK's structured fenced-response parsing and token accounting
• reject missing credentials, empty model aliases, empty responses, and malformed responses with explicit errors
• document environment-variable and structured-config usage
• add mocked, error-path, configuration, async, and opt-in live E2E tests

Usage

import os

from cltk import NLP

os.environ["LITELLM_API_KEY"] = "your-gateway-key"
os.environ["LITELLM_BASE_URL"] = "https://litellm.example.com/v1"

nlp = NLP(
    "lati1261",
    backend="litellm",
    model="your-proxy-model-alias",
)

doc = nlp.analyze(text="Gallia est omnis divisa in partes tres.")

your-proxy-model-alias is a gateway alias, not a fixed CLTK model. Different NLP instances may select different aliases, and LiteLLM can apply provider routing or fallback behavior behind each alias.

Tests

• full test suite: 62 passed, 2 skipped, 6 deselected
• LiteLLM-focused tests: 15 passed, 1 opt-in E2E skipped without live environment variables
• Ruff lint and formatting checks pass for every changed Python file
• live E2E passed against a running LiteLLM proxy with an Azure OpenAI-backed model alias, exercising the real Chat Completions response shape and CLTK fenced-block parser
• verified SDK error handling for invalid credentials, unknown models, rate limits, context-window errors, and timeouts
• verified retry exhaustion and explicit rejection of missing choices, null content, and unfenced responses

The running gateway's configured Azure Anthropic deployment currently rejects requests at the upstream account-access layer; CLTK correctly surfaces that SDK failure as OpenAIInferenceError. Provider-specific routing remains owned by LiteLLM, while CLTK consumes the same OpenAI-compatible response contract for every alias.

Compatibility

• no new required dependency; the backend reuses the existing cltk[openai] extra
• existing backends retain their current configuration paths
• API keys default to environment variables and are not logged
• LiteLLM requests are intentionally non-streaming because the existing CLTK generative pipelines parse complete structured responses