feat: InferenceParams specific for backend

[Madzionator]

This pull request refactors how inference parameters are handled for chat and agent contexts, moving from a generic InferenceParams class to a backend-specific interface and classes. Each cloud service now receives and validates its own typed inference parameters that are actually sent in API requests — but only when explicitly set by the user.

Key changes

• IBackendInferenceParams interface with BackendType Backend as the only contract — each provider defines its own fields independently
• 8 concrete param classes: LocalInferenceParams (Self), OpenAiInferenceParams, DeepSeekInferenceParams, GeminiInferenceParams, AnthropicInferenceParams, GroqCloudInferenceParams, XaiInferenceParams, OllamaInferenceParams
• Nullable properties on cloud params — null means "don't send, let the API use its own default". This avoids errors with models that reject certain parameters (e.g. OpenAI reasoning models rejecting top_p)
• Dictionary<string, object>? AdditionalParams on all param classes — catch-all for provider-specific parameters not covered by typed properties (e.g. max_completion_tokens for newer OpenAI models)
• BackendParamsFactory — creates default params for a given BackendType
• ApplyBackendParams() template method in OpenAiCompatibleService — each service override maps its typed params to API request body fields, skipping nulls
• InvalidBackendParamsException — thrown when wrong params type is passed (e.g. OpenAiInferenceParams to DeepSeek service)
• Chat.BackendParams defaults to null — ChatService auto-creates correct params via factory when not explicitly set
• InferPage wired up — InitializeChatContext() now sets WithBackend() + WithInferenceParams() based on selected backend
• Integration tests for all 7 cloud providers + local Self + param mismatch validation tests

What stays unchanged

• MemoryParams — shared across all providers (RAG/kernel memory)
• LocalInferenceParams — keeps concrete default values (llama.cpp needs them), only renamed from InferenceParams
• No UI changes beyond wiring correct params on backend switch