Add reliability testing for DSPy Adapters

[chenmoneygithub]

We are adding adapter reliability testing with the purpose of:

• Informing users of the DSPy program's quality without using DSPy optimizer.
• Compare among different adapters and vanilla LM sdk coding.
• Ensure backwards compatibility.

This initial PR only has math reasoning, and requires manual triggering via commands like:

python3 -m benchmark.adapter.math_qa.main --models openai/gpt-4o-mini openai/gpt-4o --num_threads=2 

A sample print looks like:

|                    |   Vanilla |   Vanilla with Structured Output |   DSPy (ChatAdapter) |   DSPy (JSONAdapter) |
|--------------------|-----------|----------------------------------|----------------------|----------------------|
| openai/gpt-4o-mini |        20 |                               30 |                   80 |                   80 |
| openai/gpt-4o      |        30 |                               30 |                   80 |                   70 |