feat: implement truncate_prompt endpoint

[AlexFoxalt]

fixes #28

Summary

Adds the DIAL truncate_prompt deployment sub-endpoint (POST /openai/deployments/{deployment_id}/truncate_prompt), the dry-run/planning counterpart to inline chat_completion truncation. Given a batch of chat-completion requests each with a max_prompt_tokens budget, it returns the indices of messages that would be discarded to fit — without invoking model generation (only token counting).

Implementation

• Modeled on the existing tokenize endpoint (hand-rolled FastAPI route reusing the SDK's TruncatePromptRequest/TruncatePromptResponse models).
• Reuses the app's existing truncation algorithms so dry-run results match live truncation:
  • GPT (GPT4O/GPT4O_MINI/GPT_GENERIC) → multi_modal_truncate_prompt with tiktoken.
  • vLLM (VLLM/QWEN3_ASR_VLLM) → coarse truncate_prompt with VllmTokenizer.
  • Other chat models (RESPONSES/MISTRAL/DATABRICKS/COMPLETIONS) → coarse truncate_prompt via the reused tokenize tokenizer builder.
  • Image/video/audio/anthropic deployments → 404.
• Per-input error isolation: a bad input yields a TruncatePromptError in its output slot without failing the batch; framework-level errors propagate through the existing exception handler.

Test plan

• New tests/unit_tests/test_truncate_prompt_endpoint.py: fits ([]), over-budget (system + last retained), missing max_prompt_tokens isolation, 404 for unsupported deployment, vLLM with mocked upstream tokenizer, 422 on invalid input shape.