Fix audio hash leaks, threading, and resource bugs

[eklinger]

Fixes the cluster of audio-path bugs identified in the audit: findings #7, #8, #9, #13, #14, #19, #20, #21, #29.

Correctness

• example for video test_dctvideohash.cpp not compiling #7 / Fixed video hashing and applied consistent formatting #8 Removed the entire HAVE_PTHREAD block (ph_audio_thread, ph_audio_hashes). It referenced types slice and DP and a function ph_num_threads() that were never declared anywhere in the public headers (so the block could not actually be compiled), the worker function had no return statement (UB on the void* return), and the slicing math walked one element past the input (count=10, threads=3 gave 5+3+3=11). Callers should drive ph_readaudio / ph_audiohash from their own thread pool.
• Build instructions need updating #9 ph_readaudio2 now reads src_data.output_frames_gen (the count libsamplerate actually produced) instead of output_frames (the caller-supplied capacity), eliminating the downstream over-read.
• ph_audiohash: corrected the nb_frames formula. The old expression could go negative for short inputs and was then passed to malloc.

Memory and safety

• Running make fails when building examples/test_dctvideohash.cpp #13 readaudio_mp3: every error path now closes/deletes the mpg123_handle and calls mpg123_exit. Adds NULL checks on the malloc'd output, validates channels and samplerate (channels==0 would have produced an infinite loop), clamps the per-encoding inner loop so a decode that returns fewer bytes than channels cannot over-read the decode buffer.
• Using pHash data from an older version of pHash w/the current version #14 readaudio_snd: validates channels/samplerate/frames, checks every malloc, and frees inbuf on read failure. ph_audiohash's malloc is now checked and frame-length-sized buffers are std::vectors (no more VLAs).
• Extend pHash interface with ph_mh_imagehash_from_buffer (#1) #20 / bindings / PHP build #21 Replaced double window[4096], double frame[4096], double magnF[2048], freqs/binbarks VLAs, and the per-filter new[] grid with std::vector. ph_audio_distance_ber preallocates dist once outside the per-lag loop instead of realloc'ing on every iteration; the worst-case M is just N1/block_size.
• ph_compare_blocks now returns 1.0 (worst distance) rather than dividing by zero when block_size <= 0.
• ph_audio_distance_ber and ph_audiohash zero their out parameters on every failure path; the previous code left Nc / nb_frames uninitialised when allocation failed.

Performance

• Doc for ph_hamming_distance is wrong #19 ph_bitcount uses __builtin_popcount, matching how ph_hamming_distance does its bit counting.

Header hygiene (#29)

• ph_fft.h includes <complex> (C++) instead of <complex.h> (C99) and drops the using namespace std; that was being injected into every translation unit that included it. fft() takes std::complex<double>*.
• fft() now validates that N is a power of 2 (the radix-2 recursion silently misbehaves otherwise).

Test plan

• Compiles clean (-Wall) with HAVE_AUDIO_HASH + HAVE_LIBMPG123
• FFT unit test: fft({1,4,3,2}) and fft({3,5,9,2}) produce the hand-computed DFT outputs to 1e-7; fft(N=3) is now rejected
• test_audiophash on identical 5-second sines: confidence 1.000000
• test_audiophash on 440 Hz vs 441 Hz: confidence 0.249 (low, as expected)