Overhaul lecture 21 diffusion animations: rewrite 5 scenes with matrix grids and text diffusion, fix all visual issues

- Fix GIF looping: play once and stop (gifsicle --loopcount=1)
- Fix NoiseSchedule cosine label position
- Fix ReverseProcess: black arrows, remove nn_box, keep epsilon_theta label
- Fix UNetArchitecture: side routing for bottleneck, dashed skip connections with tips, uniform arrow thickness
- Fix ScoreMatching: remove text labels, uniform arrow magnitude, expanded slide title
- Rewrite TimestepEmbedding: replace bar chart with sin/cos line plots
- Rewrite TrainingObjective: add matrix grids alongside each equation step
- Rewrite SimplifiedLoss: visual MSE pipeline (noise grids -> difference -> squared -> scalar loss)
- Rewrite SamplingProcess: replace circles with progressive denoising grids
- Rewrite DiffusionVsTransformer: text diffusion (MDLM-style mask/unmask) vs autoregressive
- Update lecture21.md: expanded ScoreMatching title, text diffusion comparison table, MDLM/LLaDA references
- Re-render all 10 GIFs and recompile slides

Author: jeremymanning