Fix test_throughput_consistency: replace np.datetime64 timing with time.perf_counter_ns

[Copilot]

test_throughput_consistency always failed in CI because np.datetime64('now') has only microsecond resolution — the timed operation completes in <1μs, so all latencies measured as 0.0, producing 0/0 = nan which fails the < 0.2 assertion.

Changes

• Timing: Replace np.datetime64('now') with time.perf_counter_ns() (nanosecond integer, no floating-point rounding)
• Sampling: Best-of-10 measurements per batch size to eliminate OS scheduling noise (standard microbenchmark practice)
• Warmup: 3 dry runs per batch size before measuring to avoid cold-cache skew
• Safety guard: Skip if min_latency < 100 ns rather than == 0 — catches unrealistically small values even with a proper timer

# Before — zero-resolution timing
start = np.datetime64('now')
_ = single_node_forward(model, x[0])
end = np.datetime64('now')
latency = (end - start).astype(np.float64) * 1e9  # always 0.0

# After — nanosecond best-of-N
samples = []
for _ in range(10):
    start = time.perf_counter_ns()
    _ = single_node_forward(model, x[0])
    samples.append(time.perf_counter_ns() - start)
best_latency = float(min(samples))