docs(moe): report shared expert banks across replicas

[kuangrenaigc-stack]

Independent MoE replicas normally allocate one full host expert-bank set per process. On a 96 GB workstation, that prevents two Qwen3.8-Flash-Next replicas from serving one 262K lane per GPU even though the GPU-side state fits.

This documentation-only change reports a tested owner/follower design that maps one immutable 63.457 GiB NVFP4 expert-bank set into two TP1 FreeToken processes. It records the protocol, normalized command, physical-sharing evidence, concurrent short and near-256K results, and the failed prefix-reuse, port-allocation, and shutdown checks. It explicitly does not present the external shim as a supported feature or tensor parallelism.

The report is intended as implementation evidence for the multi-GPU roadmap in #79. All local paths, usernames, hostnames, device identifiers, ports, workload content, and service configuration outside FreeToken were removed.

Validation:

• git diff --cached --check
• Markdown fence, tab, and trailing-whitespace checks
• privacy scan over the complete committed diff
• roadmap link returned HTTP 200
• attachment rechecked on 953565667f3141c90d0f0eb469bb2655d2407140; benchmark numbers remain attributed to the earlier tested base stated in the document