[0z5a]

Fixes #30.

rebalance_experts works with CPU float32 weights, but the greedy packing
loop currently performs individual PyTorch tensor indexing and arithmetic for
each item. Convert the sorted indices and weights to Python lists once for
this path, and use an array('f') for the pack sums so each accumulation
rounds to float32 as before. Keep the existing implementation for other dtypes
and devices. The placement policy and tie-breaking are unchanged.

Validation

• Four focused unit tests pass on the CPU of a B300 host and on a GH200 CPU.
  They compare the placement and rank outputs against the original
  tensor-scalar algorithm across random, tied, very large, and non-contiguous
  weights; exercise the hierarchical and global public policies; and check
  other dtypes, empty batches, and a non-CPU default device.
• Synthetic 61-layer, 256-logical/288-physical-expert, 16-group, 2-node,
  32-GPU simulated topology; one CPU thread, A/B/B/A order, PyTorch 2.14.1:
  on the B300 host CPU the base runs were 0.652/0.652 s and this change
  0.0439/0.0435 s; on the GH200 CPU the base runs were 0.879/0.884 s and this
  change 0.0341/0.0338 s. All three output tensors matched exactly in these
  runs. These are host-side planning timings, not distributed or model
  inference performance results.
• git diff --check and Python syntax compilation pass. No GPU execution or
  multi-node communication was required for these tests.

PRs #27 and #28 change the packing policy to address a separate placement
constraint. This change retains the current policy, and may need coordination
if either of those PRs is accepted first.