Reduce redundant video FFT work

[Josalo-III]

Checklist

• I have searched the open pull requests to confirm this change has not already been submitted.
• My branch is up to date with the target branch.
• I have tested my changes and all existing tests pass.
• Documentation is unchanged because this does not alter user-facing behavior or configuration.
• My code follows the project's coding standards (see CONTRIBUTING.md).

Description

Reduce redundant FFT work in the video demodulation path while preserving its output contract.

Motivation

The existing path computes several full-complex transforms and performs separate inverse transforms and post-transform rolls even though the dropout and video outputs are real-valued. This change avoids that redundant work.

Related Issues

N/A

Changes Made

• Use the positive-frequency half-spectrum for real-valued dropout and video outputs.
• Batch the video inverse transforms with irfft.
• Apply the existing filter sample offsets in the frequency domain.
• Add deterministic NTSC and PAL comparisons against the former full-FFT path.

Testing

• Full suite: nix develop --command python -m pytest -q — 84 passed.
• Focused tests: nix develop --command python -m pytest -q tests/test_demod_fft.py — 2 passed.
• Synthetic 200-block benchmark against current upstream main:
  • NTSC: 2.60 → 1.81 ms/block (~30.4% lower)
  • PAL: 2.83 → 2.06 ms/block (~27.2% lower)

Screenshots (if applicable)

N/A

Additional Notes

No new dependency is introduced; this uses the existing SciPy FFT implementation. The record-array output construction and public behavior remain unchanged.