fix: update imports and dependencies to address vulnerabilities

[chakravarthik27]

• Changed imports from langchain.prompts to langchain_core.prompts in prompts.py.
• Updated import from langchain.evaluation.qa to langchain_classic.evaluation.qa in helpers.py.
• Updated poetry.lock to reflect new versions of langchain and langchain-core to mitigate security issues.
• Updated pyproject.toml to require langchain version 1.3.13.

[blidiselalin]

• langchain_classic is not declared as a dependency in pyproject.toml (or poetry.lock). helpers.py imports from langchain_classic.evaluation.qa import QAEvalChain
  PretrainedModelForTranslation
• uses a hardcoded T5-style prompt for every model (translate {source} to {target}: {text}). This is incorrect for non-T5 encoder-decoder models and likely degrades quality or produces garbage for mBART, MarianMT, etc. It is also a silent behavioural change from the previous pipeline-based implementation.
• no longer validates its inputs. The constructor previously asserted isinstance(model, Pipeline); it now accepts model: Any and can be constructed with incompatible objects.

[chakravarthik27]

• langchain_classic is not declared as a dependency in pyproject.toml (or poetry.lock). helpers.py imports from langchain_classic.evaluation.qa import QAEvalChain
  PretrainedModelForTranslation

langchain_classic is from langchain package.

[pratacosmin]

Looks good to me. If this prompt format has been tested with the model, I think it's fine.

[blidiselalin]

Good to go.

[chakravarthik27]

• uses a hardcoded T5-style prompt for every model (translate {source} to {target}: {text}). This is incorrect for non-T5 encoder-decoder models and likely degrades quality or produces garbage for mBART, MarianMT, etc. It is also a silent behavioural change from the previous pipeline-based implementation.

I'll make the prompt customizable so users can refine it, with T5 remaining as the default.