fix(rocm): avoid indirect host pointers during HIP graph capture

[ErikBPF]

Summary

ROCm graph capture can fault when the fused multi-bank expert copy reads pinned-host source addresses indirectly through a device pointer table. During HIP capture only, use the existing per-bank copy path whose source is a direct tensor argument.

• eager ROCm keeps the fused path
• CUDA keeps the fused path
• only HIP graph capture falls back

Branch-contained reproducer

The public PoC branch contains the fallback, a CPU dispatch contract, and a slow HIP acceptance test using realistic Qwen3.6 geometry:

• 40 layers, 256 experts, 2,117 slots, top-k 8
• 1,179,648-byte Q4_K gate/up rows
• 720,896-byte Q5_K down rows
• 512 graph replays with changing routes and exact row fingerprints

pytest -q tests/moe/test_fused_copy.py::test_rocm_graph_capture_uses_per_bank_copy
pytest -q tests/moe/test_fused_copy.py::test_qwen36_sized_pinned_rows_survive_rocm_graph_replay

Public PoC: ErikBPF#2

Fresh public-branch evidence

A clean public-branch build on a machine with 128 GB RAM, a Ryzen 9 5950X CPU, and a Radeon RX 9070 XT 16 GB GPU (gfx1201) completed the branch's direct-copy HIP graph microbenchmark with:

• graph copies verified
• bit-exact output
• 0.216279 ms serialized median
• 0.182385 ms overlapped median

Artifact: https://github.com/ErikBPF/FreeToken/blob/draft/qwen36-rocm-full-poc/benchmarks/results/qwen36-graph-stream-overlap-gfx1201.json

This evidence validates direct-copy graph mechanics, not the exact 512-replay acceptance test and not endpoint throughput. No speedup claim is made by this PR.

Scope and overlap

Two files only. This complements #260, whose MVP intentionally disables HIP graph capture, and does not duplicate its ROCm bring-up changes. Qwen model support, FTW changes, endpoint tuning, and research documentation are excluded.