Implement Group.split() for the JACCL mesh backend

[erwinzhang7]

Addresses #3205 for the mesh case.

Only MPI implemented split, by handing the problem to MPI_Comm_split. Ring and
JACCL both threw, so there was no in-tree example of a backend building the child
itself.

A MeshGroup needs three things: the caller's rank inside the child, the RDMA
devices reaching the other members, and a side channel. The first two are arithmetic
once membership is known. The third looked like the hard part and is not:
SideChannel is defined by an all-gather function, so the parent can supply it,
which is the same shape as MPI bootstrapping a child communicator from its parent.
No new coordinator and no second discovery pass.

So split all_gathers (color, key) over the parent, selects the members sharing a
color, orders them by key and then by parent rank as MPI does, and builds the child
from the parent's own device row. MeshGroup now keeps that row, which it previously
took and discarded.

Children of different sizes stay in step because the child's all-gather is a
parent-wide collective: every rank participates in every round regardless of which
child it joined, and the length negotiation inside the parent's container all_gather
covers the whole parent at once, so two children can never disagree about how many
bytes are in flight. The closure holds a shared_ptr to the parent, so a child keeps
its parent alive.

A child can only contain directly connected members, and the hostfile is the only
description of connectivity the runtime sees, so a missing pair raises naming both
ranks rather than failing later inside the transport.

Ring reports why instead of the generic message: a subset of a ring is only a ring
when its members are contiguous in ring order, and there is no route around the gap
when they are not. So this is mesh only, deliberately.

Verified on hardware

Four M4 Pro minis wired as a full mesh, RDMA over Thunderbolt, MLX built from this
branch and installed on all four:

• splitting by parity gives children summing 5 and 10 where the parent sums 15, so
  membership and isolation both hold. Rank i contributes 2**i, so the sums name
  exactly who took part and a child leaking into the parent would return 15
• the parent still sums 15 afterwards
• a child of a child works
• an uneven 3/1 split works, including the child of size one
• a key reverses the order within a child without losing members
• a two rank child reaches 8.62 GB/s, the same as a natively configured two rank
  group, so the child uses its own RDMA connections rather than relaying through the
  parent

Depends on #4269 for the mesh case

A child can only be built from pairs the hostfile recorded. On a mesh configured as
jaccl-ring, current configure_jaccl_ring writes null for every pair that is not
a ring neighbour, so a third of the connectivity is discarded before split is ever
called. Measured on this hardware, same command, only the config code differing:
main records 8 of 12 directed pairs, #4269 records 12 of 12.

#4269 is byte identical to main on a genuine ring, so it adds where there is
something to add and changes nothing otherwise.