prov/efa: derive queued flag in efa_rdm_ope_process_queued_ope

[a-szegel]

efa_rdm_ope_process_queued_ope() took the EFA_RDM_OPE_QUEUED_* flag to process as a parameter, even though that flag is already recorded in ope->internal_flags. The four queued flags are mutually exclusive: an ope is linked onto ope_queued_list through a single queued_entry node, so at most one can be set at a time. Derive the flag from internal_flags instead of passing it, and assert the single-bit invariant.

With the flag self-derived, the progress loop in
efa_rdm_domain_progress_peers_and_queues() and the endpoint-close path collapse their per-flag probe sequences into a single call. The main loop previously issued up to four calls per queued ope -- three of which immediately returned after a redundant (internal_flags & flag) test -- each guarded by a caller-side "if (...) continue" branch. That becomes one call with one branch, removing three function calls and seven conditional branches per queued ope per progress cycle.

Add gtest coverage for the derivation: a no-op when no queued flag is set, and correct dispatch of the BEFORE_HANDSHAKE flag (which returns -FI_EAGAIN without a device round-trip when the peer has not handshaked).

[a-szegel]

I think UT's need some work, putting up draft PR to get coverage on CI