bench: allow more worker threads than cards to measure submission overlap

Two threads on one card overlap one thread's readback with the other's
dispatch, which is the bubble quantus/miner#4 and #5 target. Measuring it in
the harness first means #4 lands with a number instead of a hypothesis.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CBgs2nSi4H2mdh8kD8vMX5

crates/bench-harness/src/main.rs

@@ -40,9 +40,11 @@ struct Args {
     #[arg(long, default_value_t = 1_000_000)]
     batch_size: u32,
 
-    /// Worker threads; each is assigned its own device by the engine, so on a
-    /// multi-card host N threads measure N cards. Extra threads beyond the
-    /// card count are refused.
+    /// Worker threads. The engine assigns threads to devices round-robin, so
+    /// on a multi-card host N threads measure N cards. More threads than cards
+    /// is allowed on purpose: two threads on one card overlap one thread's
+    /// readback with the other's dispatch, which is the submission bubble
+    /// quantus/miner#4 and #5 are about, measured without touching the miner.
     #[arg(long, default_value_t = 1)]
     workers: usize,
 
@@ -307,11 +309,11 @@ fn main() {
         Arc::new(GpuEngine::try_new(args.batch_size, 0, false).expect("GPU engine init failed"));
     let devices = engine.device_count();
     if args.workers > devices {
-        eprintln!(
-            "--workers {} exceeds the {devices} GPU device(s) present",
+        log::warn!(
+            "{} workers on {devices} device(s): threads share cards, so per-worker numbers \
+             are shares and total_median_mhs is the number to compare",
             args.workers
         );
-        std::process::exit(2);
     }
     log::info!(
         "engine {} with {devices} device(s); measuring {} worker(s), {} x {}s windows after {}s warm-up, batch {}",