feat: GPU ternary decode + wire GPU into --student infer

[shift]

GPU-accelerated student inference

New: forward_decode_gpu()

Single-token decode using GPU ternary matmul with CPU KV cache:

• Q/K/V/O projections → gpu_ternary_matmul WGSL kernel
• MoE expert gate/up/down → GPU ternary matmul
• Router, attention scores, PLE, KV cache → CPU (small matrices)
• LM head → CPU (too large for iGPU 256MB buffer)

Wire GPU into cmd_infer --student

• Initialize GpuMatmulAccelerator in student inference path
• Create DispatchPlan from device profile
• generate_student() accepts GPU + dispatch params
• Decode uses forward_decode_gpu when GPU available
• Full CPU fallback for RPi and other edge devices

Remaining GPU work

• Prefill GPU path (currently CPU-only to populate KV cache)
• Upload weights to VRAM once (resident mode)
• Fix distillation double-forward (subtask 6)