Add AC3 RF audio demodulation in Python (lddecode/ac3rf.py)

[staffanu]

Summary

This is the resubmission of AC3 RF audio decoding (previously PR #1034, merged and then reverted in #1040). Following the discussion on #1034, the demodulator is now implemented in-tree in Python, following the same pattern as lddecode/efm_pll.py — numpy/scipy for filter design, numba for the DPLL and FIR kernels. There is no external repository, no nix flake dependency, and no new Python dependency (numpy/scipy/numba are already required).

With --AC3 (NTSC only), ld-decode demodulates the 2.88 MHz QPSK subcarrier and writes the raw symbols to <output>.ac3sym (one symbol per byte, values 0–3). The number of symbols demodulated during each field is recorded in the field metadata as ac3Symbols / ac3_symbols, exactly analogous to efmTValues / efm_t_values. Framing, Reed-Solomon error correction and AC3 frame assembly happen downstream in decode-orc's AC3 RF Sink stage, which already queries this metadata column; a companion decode-orc PR restores that stage's TBC-source read path.

What's in the PR

• lddecode/ac3rf.py — the demodulator: half-band Kaiser decimation stages, 2.88 MHz mixer, root-raised-cosine matched filter, and a DPLL for symbol timing recovery (numba jitclass, like EFM_PLL). Accepts arbitrary block lengths; all filter/timing state carries across calls.
• Integration in core.py replacing the dead ac3_pipe code (which shelled out to C++ tools removed in the removetools-202601 cleanup).
• tests/test_ac3rf.py — self-contained tests, no capture files needed: DQPSK loopback (clean + noisy), and an end-to-end test that synthesizes a complete NTSC LD signal (FM video carrier + AC3 subcarrier), runs ld-decode --AC3 on it, and validates the demodulated symbols and the per-field metadata counts.
• Documentation of the --AC3 behavior and the decode-orc handoff in docs/user-guide/command.md.

Validation

• Bit-exact vs the previous C++ implementation: 100% symbol agreement over three real AC3 disc captures, at 40 MHz (int16) and 62.5 MHz (uint8) sample rates.
• End-to-end on real discs: .ac3sym output decoded through the AC3 framing/RS stage produces a clean AC3 elementary stream (48 kHz, 5.1, 384 kb/s), with near-zero Reed-Solomon corrections (e.g. 3 corrected symbols in 11,952 C1 codewords), and audio verified by listening.
• Performance: ~1.6× realtime single-threaded for 40 MSps input; runs nogil so it overlaps the decoder's other threads.

Notes

• Compared to Add AC3 RF audio decoding via ac3rf Python extension #1034, the metadata records per-field symbol counts rather than offsets, matching the EFM convention and what decode-orc's metadata readers query.
• The branch currently includes the small numpy 2.0 loader fix (PR Fix .s16/.f32/.r30 loading on numpy >= 2.0 #1049) because the end-to-end test synthesizes a .s16 input; I'll rebase it out once that PR merges.

[staffanu]

Pushed ef928ce, syncing the DPLL loop gains with the reference C++ implementation (museld b82e92b): the proportional/integral gains are now computed from the second-order loop design parameters (1800 Hz natural frequency, damping 0.6, detector/VCO gain 0.3) instead of hard-coded constants. Output on real captures is unchanged (re-verified 100% symbol agreement against the updated C++ on three captures at 40 MHz and 62.5 MHz); the faster loop tracks wow/flutter substantially better per the reference implementation's testing.