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

Reproducer and evidence

Validated on Orion ROCm gfx1201 with realistic Qwen expert-row geometry:

• 40 layers, 256 experts, 2,117 slots, top-k 8
• 1,179,648-byte Q4_K gate/up rows and 720,896-byte Q5_K down rows
• 512 graph replays completed with exact fingerprints in 17.96 seconds using direct per-bank arguments
• the fused indirect source-pointer path reproduced the GPU page fault
• full endpoint completed 3 × 1,024-token generations with graph-off canary/output hashes and an empty kernel log
• later 10-sample soak remained exact and fault-free

The historical endpoint throughput varied between runs, so this PR claims graph correctness, not a speedup.

Tests

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

Packaging validation: Python compile, Ruff on the changed test, and git diff --check pass.

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, tuning, and local research docs are excluded.