refactor(blk): offload virtqueue drain to worker thread

[agicy]

Summary

This PR refactors the virtio-blk worker threading model to eliminate the intermediate producer-consumer queue between the main thread and the I/O worker thread.

Before

Guest kicks virtqueue
       ↓
  main thread (epoll)
       ↓
  parse descriptors → push to procq (TAILQ, mutex-protected)
       ↓
  worker thread → pop from procq → perform I/O → update used ring → inject IRQ

• The main thread parsed every descriptor chain and pushed blkp_req structs onto a cross-thread queue (procq).
• The worker thread popped from the queue, performed disk I/O, and injected IRQs.
• Both threads contended on procq's mutex for every single I/O request.

After

Guest kicks virtqueue
       ↓
  main thread → signal condvar (no virtqueue access)
       ↓
  worker thread → drain virtqueue directly → perform I/O → update used ring → inject IRQ

• The main thread only signals the worker via pthread_cond_signal — it never touches the virtqueue.
• The worker thread owns the virtqueue exclusively: it parses descriptors, performs I/O, adjusts the used ring, and injects IRQs, all from a single thread.
• Synchronization is reduced to one signal/wakeup per guest notification (batch), rather than per-request lock/unlock on the procq.

Performance

Test setup: ramdisk backend, fio + io_uring, 120s runtime, RK3588.

Metric	Old	New	Change
IOPS (4k rand read, qd=64)	89.0k	93.7k	+5.3%
Throughput (1M seq read, qd=16)	3831 MiB/s	4026 MiB/s	+5.1%
P99 tail latency (4k rand, qd=64)	906 us	775 us	-14.5%
P99.99 tail latency (4k rand, qd=64)	113.8 ms	76.0 ms	-33.2%

Bug Fixes

While refactoring, several pre-existing issues were fixed:

• Crash on early close: Guard pthread_join so virtio_blk_close doesn't join an uninitialized thread handle if the worker was never started.
• Incorrect written_len on read error: When preadv fails, written_len no longer carries -1 into the used-ring update (was reporting 0 bytes despite the status byte being written).
• Missing written_len for GET_ID: Now reports correct data length via the used ring so the guest sees the full device ID string.
• close(-1) on open failure: Removed spurious close(-1) in the error path when open() fails.
• Block device size detection: When fstat returns st_size = 0 (real block devices), fall back to ioctl(BLKGETSIZE64).