Let users set the amplified band in Hz (--freq-low/--freq-high)

[joeljose]

Closes #38

What changed

• New --freq-low / --freq-high (Hz, given together). They select band mode: an ideal temporal band-pass on the phase, phase += (k - 1) * bandpass(phase), so in-band motion is multiplied by k and out-of-band motion is untouched. bandpass_1d extends the series symmetrically by its own length before the FFT to avoid wrap-around. The GPU path has the same mode (_gpu_bandpass_filter, chunked, with OOM retry).
• Chose a true band-pass over mapping Hz onto the flat-top windows: the flat-top roll-off is too gradual to amplify 1.2 Hz ~10x while keeping 0.3 Hz under 2x, which is what the acceptance criterion needs. --filter {butter,...} wasn't added; one well-defined band mode is enough.
• The Parameters block always prints the band in Hz: Band: 0.8–2 Hz (ideal band-pass), or for width mode the measured half-amplitude points, e.g. Band: ~0.20–8.59 Hz (flat-top, width 80) at 30 fps (flattop_band()).
• -w still works and stays the default when no band is given (existing output unchanged). Passing it prints a deprecation note. Errors for: only one band edge, low ≥ high, -w together with a band, and a band above Nyquist (fps / 2).
• README: the Applications table now gives Hz ranges, plus the options table, an example, Temporal Filtering docs and Tips. CHANGELOG updated.

Measured gains (synthetic texture oscillation, 30 fps, --freq-low 0.8 --freq-high 2 -k 10)

Oscillation	0.3 Hz	0.8 Hz (edge)	1.2 Hz	2.0 Hz (edge)	5 Hz
Gain	1.01	6.4	9.17	4.0	1.00

The acceptance test asserts 8–12x at 1.2 Hz and < 2x at 0.3 and 5 Hz.

Tests

• Band gains (above), bandpass_1d keeps only the in-band sinusoid, flattop_band(80) ≈ 0.20 / 8.6 Hz at 30 fps.
• CLI: band run, default band printed, argument validation (4 cases), Nyquist check, -w deprecation note.
• GPU: band mode agrees with the CPU path (PSNR ≥ 60 dB). A full --gpu -k 10 --freq-low 0.8 --freq-high 2 run on face.mp4 takes 27.3 s and passes verify_output.py.
• CPU image 89 passed, 3 skipped. GPU image on an RTX 4050 112 passed, 1 skipped. GPU-path tests on CPU tensors 22 passed.