UCP/WIREUP: Reduce message size

[MattBBaker]

The addition of latency overhead values in #1544 has exceeded the size limit for UGNI wireup message sizes. This patch reduces the type from double to float, bringing the wireup message size back under the limit.

[mellanox-github]

Test PASSed.
See http://bgate.mellanox.com/jenkins/job/gh-ucx-pr/1780/ for details.

[mellanox-github]

Test PASSed.
See http://hpc-master.lab.mtl.com:8080/job/hpc-ucx-pr/3718/ for details (Mellanox internal link).

[alinask]

@yosefe @MattBBaker when packing the address, we take the value from iface_attr->latency.overhead and copy it to packed->lat_ovh: https://github.com/openucx/ucx/blob/master/src/ucp/wireup/address.c#L238

If packed->lat_ovh is now a float, shouldn't we change the source (iface_attr->latency.overhead) to the same type as well (for overflow cases)?

[MattBBaker]

It's the same way with iface_attr->bandwidth and iface_attr->overhead. The uct_iface_attr_t is huge as far as structs go so there wouldn't be much space saved. It maybe worth considering out of 'least surprise' reasons though, not surprising people by letting them think they have a double but it gets down cast into a float on the wire.