feat: support evaluating multiple input configurations per kernel task

[shangkunwang01]

This PR upgrades the JAX evaluation module to support running and benchmarking multiple input data configurations sequentially for a single kernel task.

Key Changes:

• Evaluation Data Structures (EvaluationResult):
  • Transitioned all metric fields (compiled_successfully, numerically_correct, max_abs_diff, reference_time_ms, error_trace, etc.) from scalar types to List types.
  • speedup properties now dynamically compute lists of speedups matching the input configurations.
• Harness Execution (harness_code.py):
  • get_inputs() can now return either a single tuple or a list of tuples (dynamic_args, static_args).
  • The evaluation now loops over all provided input configurations.
  • Added explicit shape matching checks to prevent jnp.allclose from crashing on shape mismatches.
• Evaluator & Utils (jax_kernel_evaluator.py, evaluation_utils.py):
  • Geometric mean speedup calculation computes the task-level speedup average.

[NinaCai]

Can you put an example (e.g. DSv4) with multiple input? So that we could follow the format for other test cases?

[shangkunwang01]

Can you put an example (e.g. DSv4) with multiple input? So that we could follow the format for other test cases?

Sure, add one commit to remove outdated files and add a dsv4 test case.

[NinaCai]

Looks like you removed a couple of ops in dataset? Is it intended?

[shangkunwang01]

Looks like you removed a couple of ops in dataset? Is it intended?

Intended. They are outdated.