First tagged release of the multi-GPU (mgpu) line, with the perf wave merged on top.

Highlights

Multi-GPU (mgpu) v0

• Monotonic-contiguous layer-placement packer (wave 1)
• Device-aware CUDA plumbing and per-device selective model cache (wave 1)
• Engine + placement scaffolding (wave 2)
• --gpu-vram / --gpu-devices wired into all four binaries (wave 2)
• CUDA-side per-device plumbing and engine-side multi-tier dispatch (wave 3a)
• v0 baseline bench harness + report (wave 3)
• Correctness tests: smoke output + deterministic token compare
• Ctx plumbed into placement, per-layer KV math, auto-mode reserve, upfront refusal

Performance wave (merged at tip)

• perf-01: prefill tensorcore path
• perf-02: opt-in split-KV / flash-decode attention kernels (with exact-window sizing fix)
• perf-04: fill SMs for routed-MoE gate/up decode via finer launch geometry
• Safer smaller q8→f16 cache reserve on high-VRAM cards

Q8_K routing

• Route q8_k experts on CPU and Metal
• Multi-block q8_k metal MoE coverage

Verifying

• Tag: `mgpu-v0.1.0`
• Commit: `0b6035b`

Getting it

```
git clone --branch mgpu-v0.1.0 --depth 1 https://github.com/cchuter/ds4.git
```
or, if already cloned:
```
git fetch --tags origin && git checkout mgpu-v0.1.0
```

Note on `main`

`main` was force-updated to this tag's commit; the previous tip is no longer reachable from `main`. If you had local work on top of the old `main`, rebase onto `mgpu-v0.1.0`.