fix(engine): resolve moe-strategy auto to fused on unified-memory GPUs (GB10)

[iamanishx]

Summary

• Detect unified-memory GPUs via cudaDevAttrIntegrated (_is_unified_memory_gpu()),
  with a FREETOKEN_UNIFIED_MEMORY=0/1 override for platforms where the attribute lies.
• On unified memory, resolve --moe-strategy auto to fused instead of the offload
  family when the model's expert format supports residency (_fused_resident_ok()),
  and skip the benchbw hybrid upgrade since CPU execution adds no bandwidth when CPU
  and GPU share one memory.
• Warn when --moe-strategy offload is explicitly requested on a unified-memory GPU.

Problem

The auto resolution always picks the offload family because on discrete GPUs a wrong
"it fits" guess is a load-time OOM. On unified-memory parts (GB10/DGX Spark, sm_121)
that rationale inverts: there is no host/device boundary, so the offload path's pinned
staging and slot gather are DRAM-to-DRAM copies, and the slot cache double-allocates
memory the GPU could read directly. Measured on GB10 in #22: offload (the default)
added a fixed 126-141 s stall to every request with ~360 s boot, while fused booted in
~50 s and served normally. Discrete-GPU behavior is unchanged: the new branch only
fires when is_integrated is set, which discrete cards never report.

Validation

Tested on a rented DGX Spark (GB10, sm_121, aarch64, driver 610.57.04, CUDA 13.0):

• torch.cuda.get_device_properties(0).is_integrated == 1 on GB10.
• pytest tests/engine tests/moe tests/kernels: 455 passed, 4 skipped.
• New tests/engine/test_moe_strategy_uma.py: 10 passed.
• Served Qwen/Qwen3.6-35B-A3B-FP8 with no --moe-strategy flag: log shows
  "Unified-memory GPU detected; auto-selecting 'fused'", 31.4 GB experts loaded
  resident, CUDA graphs captured at bs 1-2, API ready in 54 s.
• OpenAI-compatible generation verified: warm single request 15 completion tokens
  in 0.357 s; two concurrent requests in 1.86 s.

Refs: #369, #22, roadmap #79 (DGX Spark row).