Name the tensor parallel mesh dimension and slice meshes by name
`initialize_tensor_parallelism` built an unnamed 1-D mesh while the FSDP and pipeline
paths built named ones, so every consumer guarded `mesh.ndim` before slicing by name.
Name the dimension `tp` (wrapping a bare user-supplied mesh with `DeviceMesh.from_group`
so it keeps its process group) and drop the guards: the styles always receive the 1-D
`tp` mesh from `apply_tensor_parallelism`.