enh(test_common): add profiler-safe HIP-event timing path to run_perftest

[Arist12]

Problem

run_perftest times benchmark iterations by running them under torch.profiler (ROCTracer). When the benchmark is driven by an external rocprofv3 session, nesting torch.profiler inside rocprofv3 causes duplicate-flow warnings from ROCTracer and can perturb the measured kernel time. There was no way to time the same workload without entering the profiler.

The module-level import torch.profiler as tpf also caused ROCTracer to be imported eagerly on every pytest collection.

Solution

Add a FLYDSL_PERFTEST_USE_EVENTS=1 environment variable that switches the internal timing backend to paired HIP events, bypassing torch.profiler entirely. When set:

• Each of the num_iters iterations is bracketed by torch.cuda.Event(enable_timing=True) record/synchronize pairs.
• The mean over all iterations is returned as the average latency, consistent with what rocprofv3 --stats reports per dispatch.
• torch.profiler is never imported or entered.

When the variable is unset (default), behavior is identical to before.

torch.profiler is now lazy-imported in both the default timing branch and the testGraph branch, so it is only pulled in when actually used.

Typical usage

# Without external profiler — behavior unchanged, default path
python tests/kernels/test_pa_mqa_logits_fp4.py --batch 8 --ctx 131072 --next_n 1

# Under rocprofv3 — avoids nested profiler
FLYDSL_PERFTEST_USE_EVENTS=1 \
rocprofv3 --stats --kernel-trace -f csv -o /tmp/out -- \
python tests/kernels/test_pa_mqa_logits_fp4.py --batch 8 --ctx 131072 --next_n 1

Testing

• Both FLYDSL_PERFTEST_USE_EVENTS=1 and =0 paths verified to return valid (data, avg_us) tuples.
• FLYDSL_LOG_MORE=1 combined with FLYDSL_PERFTEST_USE_EVENTS=1 does not crash.
• tests/unit/test_compile_hints.py::TestCacheDisabledRegression passes (this test directly calls run_perftest).
• python -m pytest tests/unit/ -m "not l2_device and not rocm_lower": 356 passed, 2 pre-existing failures unrelated to this change.