ucp: allow eager inline sends for host memory when CUDA MDs are present

[ndg8743]

Summary

• Fixes Eager tag send inline ucp_tag_send_inline is not called on a systems with CUDA/RoCM by default #4275
• When CUDA/ROCm memory domains are loaded, ucp_proto_is_inline() conservatively disabled inline (am_short) sends for all buffers once the memory type cache became non-empty, even for host memory buffers.
• Fix adds a memtype cache lookup to positively identify host memory, allowing inline sends when the buffer is not found in the GPU memory cache.

Test plan

• Verify inline eager sends are used for small host-memory tag messages on systems with CUDA installed
• Run existing UCP tag send tests

[ndg8743]

CI failures are all infrastructure flakes on RoCE workers, unrelated to the inline send change:

• Tests roce worker 0: ud_mlx5/uct_p2p_am_test.am_short_keep_data/3 — UD endpoint unhandled timeout error (ud_ep.c:296)
• Tests roce worker 1: shm_gga/test_ucp_wireup_1sided.one_sided_wireup_rndv/5 — UMR CQ poll timeout on mlx5_1, failed to pack remote key, rndv assertion failure
• Tests roce worker 3: rc_mlx5/test_uct_perf.envelope/1 — connection timed out after ~15min
• ASAN roce worker 2: shm_ib/test_ucp_sockaddr_with_wakeup.wakeup_s2c/2 — request stuck, connection timed out, job canceled
• ASAN roce worker 0: rcx/test_ucp_sockaddr.concurrent_disconnect_s2c/10 — request stuck, connection timed out, job canceled

All are pre-existing IB/RoCE timeout patterns on swx-rain03. Retriggering CI.