[daberkow]

• A new recv() helper first tries a non-blocking dlpi::recv. When packets are already queued, which is the usual case under load, it returns one with no epoll registration and no tokio wakeup.
• When the queue is empty, it waits with the existing dlpi::recv_async, exactly as SoftNPU does today. If that wakes up and finds nothing queued, it waits again.
• tokio::task::consume_budget() lets other tasks run, since the fast path otherwise never yields to tokio.
• The two receive loops call recv() instead of recv_async().

This PR is AI assisted, please check it :)

This change along with #231 and oxidecomputer/p4#281 helped simulated omicron perf:

	Before	After
512 MiB oxide disk import	19.2 s (26.7 MiB/s)	14.0–14.5 s (35–36.5 MiB/s)
Stream-head drops during the import	2,622	1,089–1,399
rx error log lines	0	0

All in reference of #230.

[matoushybl]

matoushybl#1 I made additional changes to improve perf. Came across the same softnpu performance problems.