CI: restrict UCX transports to fix MPI init aborts on MANA runners

[mrogowski]

The weekly native workflow has been failing intermittently since late July (e.g. runs 30190900147, 30735658852, 32617182070), with a random subset of matrix jobs dying at "Test - 1 process" while the same commit passes on other
weeks. The failed jobs abort inside MPI_Init_thread, before running any test:

    UCX  ERROR mana_0: iface ... failed to create UD QP ... failed: Operation not supported
    UCX  ERROR uct_iface_open(ud_verbs/mana_0:1) failed: Input/output error
    Abort(943843343) on node 0: Fatal error in internal_Init_thread: ...
    MPIDI_UCX_init_world(260).....:
    init_worker(38)...............: ucx function returned with failed status

Some Azure runner VMs expose the MANA NIC as an ibverbs device; UCX probes it and tries to open a ud_verbs transport, but MANA does not support UD QPs, so UCX init fails and the apt MPICH (ch4:ucx) aborts. Whether a job fails depends only on which VM it lands on, which explains the random failing Python versions.

Fix: set UCX_TLS=tcp,sm,self at the job level, matching what mpi4py already uses in .ci/env.mpich. These tests run on a single host, so TCP, shared memory, and self transports are all that is needed.