Unified benchmarking harness

Benchmarking code in Iris is currently scattered across benchmark/ and examples/, with each script re-implementing the same logic (warmup loops, synchronization, timing, averaging, printing). Over time this has led to copy-pasted code, inconsistent measurement patterns, and benchmarks that are hard to reuse or automate.

It would be useful to introduce a small, shared benchmarking harness (e.g. iris.bench) that standardizes:

• warmup and iteration handling
• timing and synchronization
• basic statistics (mean / p50 / p99)
• parameter sweeps
• structured result output (e.g. JSON or dict)

This would allow both examples/ and benchmark/ to share the same timing infrastructure, while keeping example code focused on semantics rather than measurement boilerplate.

Example (sketch):

from iris.bench import benchmark

@benchmark(name="gemm_all_scatter", warmup=5, iters=50)
def run(size, world_size):
# setup tensors
# launch Iris kernel
kernel(...)

Internally you can use iris do bench and any code we have. Such a harness would significantly reduce duplicated code, improve maintainability, and make it easier to add consistent benchmarks and eventually integrate CI performance tracking.