fix: replace max_length with max_new_tokens and add basic chat context budgeting

[Shivang731]

Partially addresses #120

Summary

This PR fixes two concrete problems in the current Hugging Face generation path:

1. prompted/chat generation used max_length, which limits total tokens (input + output) and could truncate the prompt itself before generation
2. chat completions accepted full message history with no application-level trimming, making long conversations overflow context unpredictably

What changed

app/context.py

Adds a small context-management helper for the current backend path:

• ModelMetadata stores:
  • model_name
  • context_window
  • max_output_tokens
  • derived safe_input_budget
• ContextManager.trim_chat_history(...) drops older non-system messages to fit the configured input budget
• ContextManager.fits_in_budget(...) provides a simple overflow check for single prompts

app/ai.py

• initializes self.ctx in RAGAgent
• updates run_chat_completion(...) to trim chat history before rendering the chat prompt
• updates run_with_custom_prompt(...) to warn when the prompt exceeds the configured input budget
• replaces max_length with max_new_tokens in the explicit prompted/chat generation calls

app/config.py

Adds configurable settings for:

• CONTEXT_WINDOW
• MAX_OUTPUT_TOKENS

app/routes/api.py

• removes max_length from PromptedLLMRequest
• adds GET /model-info to expose:
  • model_name
  • context_window
  • max_output_tokens
  • safe_input_budget

Scope

This PR adds basic application-level safeguards for the current inference path.

It does not yet implement:

• conversation summarization/compression
• retrieval-context budgeting
• provider-agnostic runtime abstraction
• model-derived context-window discovery across backends