feat: Ternary student model inference path

[shift]

Ternary Student Inference Path

Summary

Adds --student and --moe-experts flags to the infer command for end-to-end inference through the FerrisRes student model (ternary base + optional MoE).

Usage

# Dense student (no MoE) — ternary base weights only
ferrisres infer --model-path model.safetensors --config e2b \
  --tokenizer tokenizer.json --student \
  --prompt "Hello, world"

# MoE student (2 experts) — ternary experts
ferrisres infer --model-path model.safetensors --config e2b \
  --tokenizer tokenizer.json --student --moe-experts 2 \
  --prompt "Hello, world"

Pipeline

1. Load Gemma 4 weights → gemma4_to_block_attnres() → ternary CpuBlockAttnResModel
2. Optional: dense_ffn_to_moe() → MoE expansion (2 or 4 experts)
3. generate_student() → autoregressive generation with per-step logging

Memory

All base weights are ternary {-1, 0, +1} from creation:

• Dense student: ~1.5 GB ternary (vs ~10 GB BF16 teacher)
• 2-expert MoE: ~2.1 GB ternary (vs ~17 GB FP32)
• 4-expert MoE: ~3.7 GB ternary (vs ~29 GB FP32)

New code

• generate_student(): autoregressive generator for CpuBlockAttnResModel
• Student inference path in cmd_infer() with memory diagnostics
• --student / --moe-experts CLI flags

Depends on #59 (ternary-everywhere refactor).

[3da81652]