roce_multi on a switchless ring of four: start-up warm-up and the relay #1971

Split out of #1288, whose two questions are answered and on main: per-peer devices in roce_multi (#1516) and GLM's opt-in TF_GLM_ROCE_MULTI=1 (#1524). Thanks @Bizuayeu for both and for the splits. Two pieces of work are left, both ours:

1. Start-up on a switchless ring of four. warmPeers in glm/cuda_native.zig connects every rank to every other, so a ring without opposite links fails at start. @vdalesandro hit this. Warm-up should touch only the links the ring has.
2. The relay for rings without opposite links, in roce_multi so every family on the fabric gets it, not GLM alone. Opt-in, with the wire's tests: bounded waits, the failing rank and sequence named in errors, and NCCL unless every probe matches. @Bizuayeu's single-poller and grid change (GLM-5.3-Flash at TP=3: its small all-gathers stay on NCCL, and fabric/tp's three-to-eight-rank RoCE (#1209) refuses a switchless ring with BadRecord. Which route should GLM take, and would per-peer devices be in scope? #1288, 31 Oct) can come in with the relay's rewrite if it helps there, with the relay's tests.

Worlds above two stay on NCCL by default, for the reasons in #1288.