feat: Ternary-everywhere refactor + STE shadow weight infrastructure

[shift]

Ternary-Everywhere Refactor + STE Infrastructure

Summary

Two major changes in this PR:

1. Ternary-everywhere refactor (all base weights as {-1, 0, +1})

• CpuLinear::from_weight() immediately quantizes FP32 → ternary
• CpuMoELayer stores TernaryExpert for all expert gate/up/down projections
• TernaryLinear::from_cpu_linear() uses raw ternary values (no FP32 round-trip)
• TernaryMoELayer::from_cpu_moe() copies ternary directly
• Backward pass dequantizes via .to_fp32() where needed
• Memory reduction: 2-expert MoE ~2.1 GB ternary vs ~17 GB FP32

2. STE shadow weight infrastructure (BF16-ready)

• ShadowPrecision trait: switchable BF16/FP32 for shadow weights
• BitLinear<P>: ternary base + optional shadow, STE forward, stochastic rounding, boundary noise injection, running average α
• BitMoELayer: per-block synchronized α across experts + PLE projections, ScaleSync enum (PerBlock vs PerExpert)
• Added half crate with num-traits/serde/bytemuck features
• 20 new tests (13 shadow_weights + 7 bit_moe)

3. Research papers

• ste_ternary_training.md — STE viability analysis, shadow weight memory constraints
• ternary_quality_ceiling.md — rank sweep experiment design, LoRA vs STE ceiling
• expert_count_tradeoff.md — 4 big vs 128 small experts tradeoff analysis

Test Results

• 1596 tests passing (20 new, 0 failures)
• Clean build with --features vulkan
• Pre-existing clippy warnings unchanged (516 main → 520 branch, delta from half crate)

Architecture Decisions

Decision	Choice	Rationale
Scale α	Running average (not learnable)	Prevents feedback loop death spiral
Scale sync	Per-block (all experts + PLE share α)	Prevents routing bias from scale differences
Shadow precision	BF16	50% memory savings, same dynamic range as FP32
Gradient precision	FP32	Accumulation stability

Memory Impact

Component	Before	After
Base weights (2-expert)	~17 GB FP32	~2.1 GB ternary
Base weights (4-expert)	~29 GB FP32	~3.7 GB ternary
LoRA + optimizer	~200 MB	~200 MB (unchanged)
STE shadow weights	N/A	~8.5 GB BF16 (future, not wired yet)

Next Steps (post-merge)

1. Wire ternary inference path (end-to-end test with quantized model)
2. Rank sweep experiment (ranks 4/8/16/32/64) to find quality ceiling
3. STE training pipeline (shadow weights → LoRA replacement)
4. GPU kernel for block-wide α reduction in WGSL

[3da81652]