Wallclock-fraction warmdown: cleaner LR schedule, same results

Implemented WARMDOWN_FRACTION — a wallclock-based warmdown that avoids the step_ms estimation noise in warmdown_iters.

**Problem with warmdown_iters=1200**: On 1xH100 (~1100 steps in 10 min), the warmdown effectively covers the entire training since 1200 > 1100. The early compilation steps inflate step_ms, making the LR drop to ~0.07 at step 2. It self-corrects but the LR is always in decay mode.

**WARMDOWN_FRACTION=0.3**: LR stays at 1.0 for 70% of wallclock time, then decays linearly in the last 30%. Clean, predictable, no step_ms noise.

**Results (1xH100, 10min, SwiGLU+QAT+eval@1024):**
- warmdown_iters=1200 (baseline): post-quant BPB = 1.3382
- warmdown_fraction=0.3: post-quant BPB = 1.3381

Virtually identical! The noisy warmdown_iters approach and the clean fraction approach converge to the same result. The model seems robust to the warmdown schedule details.

Now testing warmdown_fraction=0.2 (last 20%) to see if shorter warmdown helps by allowing more training at full LR.

Code change is minimal — added to Hyperparameters and lr_mul() function.