opensm/osm_multicast.c: Treat send only full members like full members on partial leave to avoid MLID leak

[myurasov-nv]

Description

osm_mgrp_add_port() counts full_members using IB_JOIN_STATE_FULL | IB_JOIN_STATE_SEND_ONLY_FULL (since 0dc7b6d, "Add support for send only full member multicast joins"), but the keep-vs-remove condition in osm_mgrp_remove_port() still tests IB_JOIN_STATE_FULL only. A port that joins with JoinState 0x9 and partially leaves with 0x1 is removed entirely while the full_members decrement is skipped: the group stays empty with full_members stuck at 1, osm_mgrp_cleanup() never runs, and the MLID is pinned until SM restart. The fix uses the same mask in the keep condition, so such a member remains as send-only-full and its final leave takes the normal remove/decrement/cleanup path.

Before	After
join(0x9) + leave(0x1): member removed, full_members stuck at 1, MLID leaked until SM restart; ~1018 cycles exhaust the MLID space and every new group creation fails with ERR 1B19 (SA status NO_RESOURCES) despite no active members	member kept as send-only-full; the final leave reclaims the MLID; 1100 join/partial-leave/final-leave cycles complete with no failures, plain full join/leave behavior unchanged

Reproduction

1. Build OpenSM from master and run it against an ibsim simulated fabric (2 switches, 4 HCAs) with the umad2sim shim: LD_PRELOAD=umad2sim.so opensm -f opensm.log.
2. From a umad SA client, send an MCMemberRecord SubnAdmSet join with JoinState 0x9 for a fresh MGID, then a SubnAdmDelete leave with JoinState 0x1; repeat with a new MGID per cycle.
3. Unpatched: after ~1018 cycles every further join fails with SA status 0x0100 (NO_RESOURCES) and the log shows mcmr_rcv_create_new_mgrp: ERR 1B19: get_new_mlid failed request mlid 0x0000, while SA MCMemberRecord queries confirm the pinned groups have no members. Only an SM restart recovers. Patched: the same loop runs 1100 cycles clean and each final leave reclaims its MLID.