KV cache post: cut to ~8min, drop math/jargon, classroom analogy from top

Big simplification pass for non-ML readers:
- Cut: embeddings-as-vectors section, full attention formula, multi-head
  matrix details, transformer block decomposition (LayerNorm/FFN/residual),
  causal masking matrix diagram, naive cost big-O analysis, full memory
  math breakdown by dimensions.
- Kept: tokenization (one paragraph), classroom + Q/K/V cards, layers and
  parallel classrooms (brief), generation loop, KV cache (with SVG diagram),
  memory cost (with calculator widget), modern serving optimizations as
  folder operations, one-breath summary.
- Read time: 14 min → 8 min.

Author: CodeYogiCo