Issue #64: simple_backward fix - split in/out offsets #519

Task

Fix simple_backward to use separate in/out offsets (not shared) for correct gradient flow.

Acceptance Criteria

• Split offsets_in and offsets_out in simple_backward
• Correct gradient accumulation: W.grad += grad_out - grad_in
• No gradient sharing between input and output

Test

• overfit-100: 500 steps, BPB < 0.5
• Run in experiments/backward/overfit_100/

Refs

• EXP-001: Trinity 3³ physics
• EXP-010: Ternary block alignment

Implementation

Current: shared offsets cause gradient mixing
Fix: separate offsets_in (from tokens) and offsets_out (to tokens)

Commit

"fix(backward): split in/out offsets in simple_backward"
Refs: EXP-001, EXP-010