e2e: synchronous rocm calls in async steps freeze every other scenario #720

cucumber polls every in-flight scenario from a single task, and most steps run the rocm binary synchronously (run_rocm and its siblings, about 180 call sites). While one scenario's command runs, no other scenario is polled at all. Any step with a time window can then fail for reasons that have nothing to do with the product.

Recent cases:

• E2E chat-06 flakes on the mock lane: /v1/models times out against the loopback mock #661/fix(e2e): stop executor stalls from timing out chat-06 on the mock lane #662: the chat-06 request timeout counted time spent waiting for the harness.
• test(e2e): don't count harness stalls against the quit timeout #719: dash-09 and dash-34. The quit step's 500 ms wait came back after 102 s while other scenarios ran installs back to back, so the 30 s deadline passed before the driver could answer the prompt. Main's CI went red after fix(therock): pin torchaudio to the requested ROCm 10 build #630 added four scenarios to the mock lane.

Each of those fixes compensates in one step: #662 runs the request on its own task, and #719 extends the deadline by the time the driver was stalled. The root fix is to make the rocm-running helpers async (for example tokio::process, or spawn_blocking plus .await), so a long command in one scenario no longer stops the others.