libnvme: rewrite the discovery walk as a depth-capped, deduped walker

[martin-belanger]

Rewrites libnvme's discovery walk (_nvmf_discover()) from scratch, replacing the closed #3828, and breaks it into smaller, logical helpers to make the flow easier to follow.

Four main things in here:

1. --force renamed to --no-reuse (kept as a deprecated alias): --force used to double as both "force persistence" and "skip reusing an existing connection"; now that --persistent=force exists as its own explicit option, --force only means the latter, so its name no longer matches what's left of its job. This is also the widest-reaching commit -- the only one that touches CLI docs, generated accessor bindings, and src/fabrics.c, not just libnvme/src/nvme/fabrics.c.
2. NBFT discovery now walks through the same code nvme discover/connect-all use, instead of a second, forked implementation of the same walk.
3. A new TID-keyed visited list, so the Discovery Log Page walk tracks every Discovery Controller it has already visited in this walk -- closes a gap where a referral graph with more than one path to the same DC, or an outright cycle, could be walked more than once.
4. Several real bugs, found and fixed along the way (list below).

Bugs found and fixed along the way

None of these were part of the original plan; all found while working through the surrounding code.

• Three fctx->hooks.* calls (connected, already_connected, decide_retry) had no NULL check, even though every hook is documented-optional and every other call site in the file guards them. An external caller passing NULL for a hook it doesn't need would crash.
• A spec-legal all-zero DCNQNHOR (NVMe Boot Specification rev 1.4: "no unique NQN, use the well-known Discovery NQN") was treated as a hard parse failure, silently dropping the whole Discovery Descriptor -- including the common case, since most descriptors have no unique NQN at all.
• Once that parsing bug was fixed, libnvmf_discover_nbft() still never read the parsed NQN when present, always connecting with the well-known one instead.
• When --device didn't resolve to a real controller, fctx->persistent got silently forced to NO, discarding whatever the user actually asked for with --persistent.
• An already-connected referral was skipped entirely, without ever walking its own referrals.
• A discovery self entry was identified by pointer identity (cl == c) instead of its spec-defined SUBTYPE.
• A referral's KATO was requested after the connect already happened, too late to take effect.
• hooks.connected never fired for anything discovered and connected during the walk itself -- connect-all never reported per-device progress for its own discovery-driven connects, even though the hook exists and fires correctly everywhere else.
• NBFT: a Discovery Descriptor's host_iface was freed before its own walk read it from every connect made during that walk (moot now -- ctrl_params_dup()'s __cleanup_ctrl_params already prevents this by construction).
• NBFT: the Discovery Descriptor connection's own DHCP retry cleared the destination address (traddr) instead of the local one (host_traddr) -- a retry that could never succeed.

Commits

1. guard three unchecked fabrics hook calls against NULL -- every hook is optional; three call sites weren't checking.
2. rename discover/connect-all --force to --no-reuse -- --force no longer means what its name says now that --persistent=force exists separately; --force kept as a deprecated alias.
3. use a Discovery Descriptor's own NQN when specified -- a spec-legal all-zero DCNQNHOR was mishandled as a hard parse failure, dropping the whole descriptor.
4. unify discovery ctrl lookup-or-create into dc_open() -- one path for "reuse or create a discovery controller," used by discover, connect-all, and (later) NBFT.
5. rewrite _nvmf_discover() as a depth-capped, deduped walk -- the core change. Fixes four real bugs: an already-connected referral was skipped without ever walking its own referrals, a self entry was misidentified by pointer identity instead of SUBTYPE, KATO was requested too late to take effect, and hooks.connected never fired for controllers discovered and connected during the walk. Adds a referral depth cap and a TID-based visited set for cycle detection. dc_decide() is a new pure, unit-tested function (9 cases in test-fabrics.c) driving every disconnect decision.
6. fold NBFT discovery onto the shared walker -- nbft_discovery() was a forked copy of the walk logic, missing the depth cap and visited-set the general path just gained. Replaced with the same dc_open()/dc_walk() machinery via a new private connect_leaf hook, fixing two more pre-existing bugs along the way (see commit message).

Testing

• meson compile clean under default, -Dwerror=true, and -Dfabrics=disabled.
• meson test passes, excluding the pre-existing, unrelated fabrics-mock-cli LD_PRELOAD failure that's also present on a clean master.
• New test_dc_decide() unit test exercises dc_decide()'s disconnect logic directly.

[igaw]

Generally looks good, just a few nitpicks.

The only big question is on the interpretation what 'referral entries deeper than 8 levels' means.

[igaw]

Thanks!