fix: handle missing Jinja2 variables gracefully in LLM node

[EvanYao826]

Summary

Fixes langgenius/dify#38655

When an LLM node uses Jinja2 templates with variables that may not exist in the variable pool (e.g., because they come from a branch path that wasn't taken during workflow execution), _fetch_jinja_inputs should return empty strings for missing variables rather than raising VariableNotFoundError.

Changes

• src/graphon/nodes/llm/node.py: Changed _fetch_jinja_inputs to use variable_pool.get() directly instead of _get_required_variable(). When a variable is not found in the pool, it defaults to an empty string instead of throwing VariableNotFoundError.

Rationale

This is consistent with how _render_jinja2_message (in the same file) already handles missing variables: variable.to_object() if variable else "". The Jinja2 template can use {% if var is defined %} to handle optional variables, but that check never gets reached because the code fails before rendering.

[QuantumGhost]

This PR seems related to and potentially conflict with #214. What's the motivation for these two PR @EvanYao826 ?