feat: add LiteLLM as unified LLM provider

[RheagalFire]

Summary

• Adds LiteLLM as a model provider, giving nexent users access to 100+ LLM providers (OpenAI, Anthropic, Google Gemini, Azure, Bedrock, Ollama, etc.) through litellm.completion() as an SDK dependency.
• Two integration points: backend provider for model discovery + SDK model for inference.

Changes

• backend/consts/provider.py: added LITELLM = "litellm" to ProviderEnum
• backend/services/providers/litellm_provider.py: new LiteLLMModelProvider for model discovery via /v1/models
• backend/services/model_provider_service.py: wired LiteLLM into get_provider_models() dispatch
• sdk/nexent/core/models/litellm_llm.py: new LiteLLMModel with streaming, token tracking, tool calling, drop_params=True
• sdk/nexent/core/models/__init__.py: registered LiteLLMModel
• test/sdk/test_litellm_model.py: 17 tests (init, streaming, credentials, edge cases, live E2E)

Tests

Unit tests (17/17 pass):

TestInit::test_basic PASSED
TestInit::test_with_credentials PASSED
TestCall::test_streaming_output PASSED
TestCall::test_api_key_forwarded PASSED
TestCall::test_api_key_omitted_when_none PASSED
TestCall::test_api_base_forwarded PASSED
TestCall::test_drop_params_set PASSED
TestCall::test_response_format PASSED
TestCall::test_stop_sequences PASSED
TestCall::test_token_tracking PASSED
TestEdgeCases::test_empty_stream PASSED
TestEdgeCases::test_chunk_without_choices PASSED
TestEdgeCases::test_context_length_exceeded PASSED
TestEdgeCases::test_auth_error_propagates PASSED
TestEdgeCases::test_import_error PASSED
TestEdgeCases::test_stop_event_interrupts PASSED
TestLiveE2E::test_live_streaming PASSED
============================== 17 passed in 3.64s ==============================

Live E2E (Anthropic claude-sonnet-4-6 via Azure Foundry, streaming):

Live E2E response: 'OK'

Risk / Compatibility

• Additive only. Existing providers and OpenAIModel untouched.
• litellm is lazy-imported so the base install is unaffected.
• drop_params=True ensures cross-provider compatibility.

Example usage

from nexent.core.models import LiteLLMModel

# Any LiteLLM-supported provider
model = LiteLLMModel(
    model_id="anthropic/claude-sonnet-4-20250514",
    api_key="sk-ant-...",
)

# Or Google Gemini
model = LiteLLMModel(model_id="gemini/gemini-2.5-flash")
# export GEMINI_API_KEY=...

response = model([{"role": "user", "content": "Hello!"}])
print(response.content)

[RheagalFire]

cc @Dallas98 @WMC001

[JasonW404]

安全风险：verify=False 禁用了 SSL 证书验证，生产环境下容易遭受 MITM 攻击。项目已有 ssl_verify 配置字段，应该从 provider_config 读取并默认为 True。

[JasonW404]

所有通过 LiteLLM 发现的模型都使用同一个 DEFAULT_LLM_MAX_TOKENS，不区分实际能力。gpt-4o 有 128K context，gpt-3.5-turbo 只有 16K。这会与 PR #3293 的 capacity management 系统冲突。建议从 LiteLLM 的 model cost map 查询实际能力，或允许 operator 覆盖。

[JasonW404]

ImportError 在 __call__ 时才抛出，意味着错误只在第一次调用时才暴露。建议在 __init__ 时就验证 litellm 是否可用，实现更早的失败反馈。

[YehongPan]

[代码规范] except Exception: 过于宽泛，建议捕获更具体的异常类型，避免掩盖潜在错误。

[YehongPan]

[代码规范] except Exception: 过于宽泛，建议捕获更具体的异常类型，避免掩盖潜在错误。

[YehongPan]

[代码规范] except Exception: 过于宽泛，建议捕获更具体的异常类型，避免掩盖潜在错误。

[WMC001]

LGTM. LiteLLM integration follows the existing provider pattern. No issues found.