[PCIe] add group-consistent i8_hier DCP transport

[yatesdr]

What

Adds i8_hier, a four-rank INT8 PCIe all-reduce mode for systems arranged as two fast GPU pairs with a slower inter-pair path.

• the first reduce-scatter and final all-gather remain inside each fast pair;
• only the pair-owned half of the compressed tensor crosses the slow boundary;
• matching owners combine pair partials in a canonical order;
• every rank reconstructs the final BF16 result from the same final INT8 payload;
• unsupported world sizes fail closed.

The implementation uses the PCIe lifecycle supplied by upstream SparkInfer PR local-inference-lab#105. It does not add a rank-local fallback or alter PR local-inference-lab#105's initialization, replay, header-packaging, or teardown contracts.

Why

On four-GPU PCIe systems with a 2+2 topology, a flat ring sends more data across the slow boundary than necessary. i8_hier keeps the existing rank-consistent block-INT8 codec while matching the collective schedule to that topology.

Validation

• includes focused mode-normalization, rank-consistency, and GPU collective coverage;
• the exact files in this branch are byte-identical to the qualified r17 image;
• the TP4/DCP4 production qualification selected B12X PCIe collectives on every rank with no PyNCCL fallback, collective divergence, capture failure, or restart;
• cold 350k retrieval returned the exact expected value 738216;
• MTP3/C16 completed at 234.05 aggregate tok/s with 16/16 requests active and zero errors.

Dependency

This draft is stacked on local-inference-lab#105. Its base branch is an immutable copy of that PR head so the review diff contains only i8_hier. Retarget to master after local-inference-lab#105 merges.

Review stack: #1