Expose real NBD read and cache metrics

[RenanGBarreto]

Problem

/api/status currently reports requests_total, bytes_served, and read_errors as zero even after clients read from the virtual disk. These fields are defaults, not live nbdkit/plugin metrics.

This prevents diagnosing slow directory enumeration, file downloads, cache behavior, and read failures.

Plan

1. Add thread-safe counters in the nbdkit plugin for requests, bytes, and read errors.
2. Expose counters through the service status model.
3. Define counter reset and generation semantics.
4. Add API regression tests proving counters reflect reads and errors.
5. Use metrics during SBC performance verification.