feat: multiple GPUs support and async dashboard spinner

[gmandelli]

Summary

Owlrun previously assumed a single GPU end-to-end. Detectors only parsed the first nvidia-smi line / first WMI result, model fit math used a single VRAM number, and the Ollama subprocess
was launched with invalid CUDA_VISIBLE_DEVICES=all syntax and no multi-GPU scheduling hints. This PR makes multi-GPU rigs first-class on Linux, Windows, and macOS (eGPU).

The gateway registration payload is intentionally not modified.

Not Tested: Windows, Apple Silicon

Changes

Detection (internal/gpu)

• New GPUDetail struct and Info.GPUs []GPUDetail. VRAMTotalMB is now documented as the SUM across cards.
• New helpers Info.LargestSingleVRAMGB() and EffectiveVRAMGB(info, split).
• New shared parseNvidiaSmi() (in nvidia_parse.go) that enumerates every GPU. Used by Linux, Windows, and macOS NVIDIA detectors.
• Windows AMD detector switched to ConvertTo-Json -AsArray, dropped Select-Object -First 1, with a PowerShell 5 single-object fallback.
• Apple Silicon path now populates GPUs[] with its single entry for consistency.

Ollama subprocess env (internal/inference)

• New shared gpu_env.go builds platform-agnostic GPU env vars:
  • OLLAMA_NUM_GPU=99
  • OLLAMA_FLASH_ATTENTION=1 (NVIDIA) / 0 (AMD)
  • OLLAMA_SCHED_SPREAD=1 only when split is enabled AND Count > 1
  • Explicit device index lists (0,1,...,N-1) for CUDA_VISIBLE_DEVICES, HIP_VISIBLE_DEVICES, ROCR_VISIBLE_DEVICES
• Removed the invalid CUDA_VISIBLE_DEVICES=all literal.
• ollamaEnv signature is now (info, gpuSplit) on all three platforms. Linux and macOS previously returned os.Environ() unchanged — they now actually emit GPU env vars.
• Manager gained gpuSplit field, SetGPUSplit/GPUSplit accessors, passes the flag into ollamaEnv on subprocess start.

Config (internal/config)

• New InferenceConfig.GPUSplit bool, loaded from [inference] gpu_split (default false).
• New SaveGPUSplit(on) helper.

Dashboard (internal/dashboard)

• Status.GPU gained Count, Split, and GPUs []GPUEntry for the per-card breakdown.
• New SetGPUSplitFunc setter, /api/set-gpu-split route, handleSetGPUSplit.
• New HTML toggle #gpu-split-row (hidden via JS unless gpu.count > 1), JS function setGPUSplit, and a status-sync hint that displays the actual largest-single vs pooled VRAM numbers.

Tray wiring (tray.go, tray_linux.go, tray_other.go)

• Status fill populates Count, Split, and GPUs[].
• All model-selection callers now use gpu.EffectiveVRAMGB(info, cfg.Inference.GPUSplit). Default behaviour is largest single GPU — safe upper bound for models that don't shard across
  cards.
• After Manager creation, SetGPUSplit(cfg.Inference.GPUSplit) is invoked so the initial Ollama launch uses the right env.
• New SetGPUSplitSetter saves config, updates the manager, restarts Ollama, re-selects models with the new effective VRAM, reloads the primary, and reconnects to the gateway. Rejects the
  call when Count < 2.

Installer (installer/install.sh, installer/install.ps1)

• Both installers enumerate every GPU instead of stopping at the first.
• Step 1 prints a per-GPU breakdown when Count > 1.

Behaviour changes for users

• Single-GPU users: no functional change. The dashboard split toggle is hidden.
• Multi-GPU users: model fit defaults to the largest single card (previously summed all VRAM, which was overly optimistic). A new dashboard toggle ("Split model across GPUs") opts into
  pooled VRAM with OLLAMA_SCHED_SPREAD=1 and triggers an Ollama restart + model reselection.

Async operation feedback

• New server-published busy field on /api/status for long-running ops.
  handleSetGPUSplit and handleSwitchModel set it for the duration of the
  call.
• Dashboard shows a fixed top banner ("Applying GPU split…", "Loading model X…")
  with a spinner and disables buttons/checkboxes while busy. Optimistic update
  on click; server state takes over on the next poll.

Model download visualization fix

• Server now publishes pulling_model (the tag currently being downloaded)
  alongside pulling.
• Dashboard model list previously fell through to a generic spinner branch for
  every model whenever any pull was active. Fixed: only the model matching
  pulling_model shows the "Downloading…" spinner. Other installed models keep
  their Activate/Remove buttons; non-installed models render a disabled
  Download button until the in-flight pull finishes.

What is NOT changed

• internal/marketplace/gateway.go and registerPayload are untouched (per request).

Test plan

• make build-all (Linux, Windows, macOS amd64/arm64) passed
• go test ./internal/gpu/... ./internal/inference/... ./internal/dashboard/... ./internal/config/... passed
• make test (full suite) passed
• Manual: 2x NVIDIA rig — verify dashboard shows both cards, toggle restarts Ollama, OLLAMA_SCHED_SPREAD=1 is set when on passed
• Manual: 2x AMD rig on Windows — verify WMI enumeration returns both cards not tested
• Manual: single-GPU rig — verify toggle row stays hidden and behaviour is unchanged passed
• Manual: Apple Silicon — verify no regressions not tested