[v2.4.x] prov/efa: use dmabuf for CUDA MR registration

[sara4dev]

Summary

• include CUDA in the EFA provider's implicit dmabuf MR registration path
• use the existing ofi_hmem_get_dmabuf_fd() and ibv_reg_dmabuf_mr() flow for FI_HMEM_CUDA
• remove the stale TODO that said CUDA still needed this fallback

Root Cause

CUDA HMEM registrations without FI_MR_DMABUF were not included in the EFA provider's dmabuf path, unlike Neuron and ROCr. That made CUDA memory fall through to plain ibv_reg_mr() with a GPU virtual address, which can fail with EFAULT.

References #12019.

Validation

Tested in a GB200 cluster

[a-szegel]

Can you please sign your commit? git commit --amend -s

[shijin-aws]

@sara4dev These fixes are already part of v2.5.x as part of c9e3c0c, would u mind trying it ? or you prefer to stay on v2.4.x