Pack parallel gradient allreduces

[XingyuZhang2018]

Summary

• Add a packed allreduce helper that groups gradient arrays by eltype/backend and reduces each group through one allreduce_p2p! call.
• Use the packed helper in parallel's reverse pass for non-split argument gradients, preserving dense-gradient semantics while reducing Allreduce launch count.
• Add optional MPI coverage for packed allreduce correctness.

Test Plan

• packed allreduce 2-rank PASS via MPI.mpiexec() -n 2
• optional MPI single-process test file: 55/55 pass
• FLmap_parallel serial-vs-parallel gradient check: relative errors (0.0, 0.0, 0.0, 0.0)
• manual one-step C4v VUMPS trace: PASS