fix(evaluator): inject configured LLM into custom metrics

[thesamet]

Summary

Custom metrics previously loaded the LLM themselves by reading a hardcoded config.yaml path at import time, causing an immediate crash when the configured provider differed from what the metric expected (e.g. running with Anthropic while metrics tried to instantiate an OpenAI client).

• The evaluator now inspects each custom metric's __init__ signature and passes the already-configured LLM instance when an llm parameter is declared
• Metrics without the parameter continue to work unchanged (fully backward-compatible)
• Accessing self.llm when no LLM was injected raises a descriptive RuntimeError instead of an opaque AttributeError
• Updated both built-in example metrics (bank-insurance, e-commerce) to use the injected LLM
• Updated docs to show the llm=None injection pattern

Closes #131

Changes

• arksim/evaluator/evaluator.py: detect llm parameter in custom metric constructors and inject the configured LLM instance
• arksim/evaluator/base_metric.py: expose llm as a property on both base classes that raises a clear RuntimeError when accessed without injection; add llm parameter docs to both QuantitativeMetric and QualitativeMetric docstrings
• examples/bank-insurance/custom_metrics.py, examples/e-commerce/custom_metrics.py: remove hardcoded config loading, accept injected LLM via self.llm
• docs/main/evaluate-conversation.mdx: update both metric type examples to show the llm=None injection pattern
• tests/unit/test_evaluator_class.py: new tests covering injection for quant and qual metrics, backward-compat (no-param) case, and RuntimeError when self.llm is accessed without injection

Documentation

• Updated relevant docs in docs/ (if behavior, config, or API changed)
• Updated README.md (if installation, quickstart, or usage changed)
• No docs needed (explain why below)

How to Test

• ruff check . passes
• ruff format --check . passes
• pytest tests/ passes
• Manual verification: run evaluation with Anthropic config against example metrics -- no crash, LLM injected correctly

Notes

Backward-compatible: any existing custom metric that does not declare an llm parameter in __init__ is instantiated as before. Only metrics that opt in by declaring the parameter receive the injected LLM. Metrics that declare llm but try to use self.llm without injection now get a descriptive error pointing to the fix.

Reviewers

/cc @arklexai/arksim-maintainers