feat(dsp): non-uniform partitioned convolution — IR cap 2 s → 10 s

[ealtun21]

Closes #41.

What

Two-stage Gardner partitioning in the convolution engine:

• Head — unchanged uniform 256-sample partitions covering the first 8192 taps: latency stays exactly BLOCK (256 frames).
• Tail — 4096-sample partitions (FFT 8192) for everything past 8192 taps. HEAD_LEN = 2 × TAIL_BLOCK is load-bearing: a completed tail block finishes one full 16-sub-block cycle before its earliest output is due, so the tail FFT + frequency-delay-line MAC are spread across the cycle's sub-blocks — no CPU spike, no worker thread, deterministic allocation-free RT path (audiodg-safe).
• MAX_IR_SECONDS 2.0 → 10.0. Short IRs (≤ 8192 taps) allocate no tail state and behave byte-for-byte as before.

Public API, IPC protocol, APO IR-blob format and latency are unchanged.

Measured (release, per 256-frame stereo call)

IR	before	after
2 s @48 kHz	~1 ms (750 partitions)	54 µs
10 s @48 kHz	rejected	165 µs
10 s @96 kHz worst-case sub-block	—	492 µs vs 2.67 ms budget

Memory note: FDL + kernel ≈ 32 B/tap/ch each (10 s stereo @96 kHz ≈ 90 MB); a realfft follow-up can halve it.

Tests

• 6 new DSP tests: sparse-delta long IR vs analytic reference across the head/tail boundary, 10 k-tap dense IR vs direct convolution, long-IR block-size invariance, 3 s no-longer-truncated, tail-state reset on re-enable, rebind + channel-resize with a long IR. All 16 existing convolution tests unchanged and green.
• New daemon IPC test: 100 001-tap WAV loads end-to-end, tail tap convolves at the exact frame.
• benches/chain.rs: convolution bench at 0.1 s / 2 s / 10 s.
• Ignored manual RT-budget test (worst_case_block_time_fits_rt_budget).

Verified on

• Linux: make check green (fmt, clippy -D warnings, full workspace suite).
• macOS (arm64, 26.2): full workspace suite green.
• Windows (Win10 VM): dsp 145 + apo 27 green; live in audiodg — 144 001-tap IR accepted (+5.3 ms reported), identity 3 s IR bit-transparent vs baseline, and a 0.5-amplitude echo at 2.5 s (past the old cap, deep in the tail stage) measured in WASAPI loopback at exactly the predicted comb: 700 Hz +3.52 dB, 900.2 Hz −6.02 dB.
• Linux live long-IR run deferred to the scheduled audio-cut window (live daemon must not be restarted until then).