Explain the method from first principles (docs/theory.md)

[joeljose]

Closes #74

New: docs/theory.md (≈3,500 words)

A tutorial from first principles, linked from a new README section, "How to learn this":

1. What magnifying motion means.
2. Eulerian vs phase-based, including Wu et al.'s (1+α)δ < λ/8 bound.
3. Local phase: the derivation that a shift δ changes phase by ωδ, why k·Δφ moves an edge kδ, and the wrap-around limit.
4. The DTCWT: why not the DWT, the two trees and Hilbert pair, level-1 vs Q-shift filters (every --biort/--qshift option with its tap counts), six orientations, levels/sizes/4:1 redundancy, what --nlevels means spatially.
5. Following phase over time: normalise, conjugate product, cumulative sum, and why zero coefficients give no motion.
6. Temporal filtering: width mode as G(f) = H2(f)·[1 + (k−1)(1 − Hlow(f))], zero-phase windows, the 0.2327 factor, band mode, frequency resolution.
7. Reconstruction.
8. The unmagnified lowpass residual.
9. Validation summary.
10. Colour (yiq).
11. Choosing parameters.
12. Comparison table: Eulerian, steerable, Riesz and DTCWT.
13. Glossary.
14. Nine references.

New: scripts/make_theory_figures.py

Builds the five figures in docs/images/theory/ (536 KB total) from the actual pipeline functions. Needs matplotlib, which the script's docstring says; it isn't a project dependency.

• DTCWT sub-bands of a face crop (3 levels × 6 orientations).
• Phase change against sub-pixel image shift per level. Measured slopes give effective wavelengths of 4.1 / 6.0 / 11.2 / 19.6 / 43.3 px for levels 1–5. Half a turn is reached at 2.1 / 3.0 / 5.6 / 9.8 / 21.7 px, which explains the ~3 px limit measured in Reduce artefacts at high k: amplitude-weighted phase smoothing, limits on large phase shifts, per-level magnification #39.
• Shift invariance: a blob moved in 1/8 px steps. DTCWT level energy stays within 1%; Haar DWT swings 54–62%.
• Width-mode filter responses and the resulting gain against band mode.
• One face.mp4 coefficient through φ, φ0, detail and the ×10 detail.

README corrections

• The pipeline diagram used "complex division" (it has been a conjugate multiply since CPU path: exact-zero wavelet coefficients produce NaN, which shows up as black regions in the output #22). It now also shows band mode, and the text mentions --color-space yiq.
• Unsourced claims replaced:
  • "10–100x amplification" → Wadhwa et al.'s reported ~4× over linear Eulerian, plus our measured ~3 px limit.
  • "~21x overcomplete" → configuration-dependent.
  • "~5x faster than steerable pyramids" → removed; the table now shows 4:1 redundancy.
  • GPU "~5x faster" → the measured face.mp4 times (24 s vs 60 s).
• Project structure and CHANGELOG updated.

I left out statements I couldn't verify, e.g. whether Wadhwa et al. magnify luma only. The example coefficient's oscillation is described as measured: about 0.3 Hz, slow head motion rather than the pulse.