Avoid mutating LLM prompt dataframes

[pragnyanramtha]

Summary

• build LLM prompt template completions without writing an internal __mdb_prompt column into the caller DataFrame
• return a typed empty integer array for no-placeholder non-strict prompt templates
• add a regression test that preserves a user-provided __mdb_prompt column

Why

get_completed_prompts is used by LLM/model handlers to render prompt templates. The helper previously mutated its input DataFrame by adding __mdb_prompt, which could leak internal state into downstream code or overwrite a real user column with the same name.

Validation

• .test-venv/bin/python -m pytest tests/unit/various/test_llm_utils.py -q -> 2 passed in 13.09s
• /tmp/mindsdb-ruff-venv/bin/ruff check mindsdb/integrations/libs/llm/utils.py tests/unit/various/test_llm_utils.py -> All checks passed!

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[pragnyanramtha]

I have read the CLA Document and I hereby sign the CLA

[pragnyanramtha]

recheck