Downgrade MULTI sentinel in rev_host_clear

[ChuWeiChang]

As stated in #38 (comment), MULTI sentinel in host_to_vfd[] is never recover to vfd once it is set. This PR fixes it by adding a per-host-FD reference count host_fd_refs[] that tracks how many VFDs alias each host FD:

rev_host_set increments the count before writing the slot.
rev_host_clear decrements it on close. When the count falls to 1, it searches for the surviving VFD and writes it directly into host_to_vfd, downgrading MULTI back to a valid mapping. When the count reaches 0, the slot is cleared to KBOX_HOST_VFD_NONE as before.

A new unit test test-fd-table-refcount.c covers the single-holder, two-sharer, and downgrade cases.

---

Summary by cubic

Fixes stale MULTI entries in the reverse host-FD map by tracking per-host FD refs and downgrading back to a direct mapping when only one VFD remains. Restores correct reverse lookups after shared FDs are closed.

• Bug Fixes
  • Add host_fd_refs[]; initialize in kbox_fd_table_init.
  • rev_host_set increments; rev_host_clear decrements.
  • On refs == 1, search for the survivor and restore a direct mapping (if the search hits the closing VFD, hide it and retry). On refs == 0, clear to KBOX_HOST_VFD_NONE.
  • Add unit test (test-fd-table-refcount) covering single, shared, and downgrade cases.

^{Written for commit 0000cbd. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 5 files

[ChuWeiChang]

net-wget-external is failing due to DNS resolution of www.google.com timing out in the runner — this is unrelated to the fd-table changes in this PR. Could @jserv re-run the integration job?

[jserv]

net-wget-external is failing due to DNS resolution of www.google.com timing out in the runner — this is unrelated to the fd-table changes in this PR. Could @jserv re-run the integration job?

DONE. Make test script more robust.