A real, deterministic AnalyserNode FFT

The Web Audio AnalyserNode - what visualisers use to draw a spectrum or a waveform - was a pass-through: it summed its inputs and reported nothing. It is real now. During an offline render the engine captures each analyser's most recent fftSize samples and runs a genuine FFT over them, so getFloatFrequencyData / getByteFrequencyData return an actual magnitude spectrum (Blackman window, dB scale, min/max/smoothing honoured) and getFloatTimeDomainData / getByteTimeDomainData return the real waveform. A subtle correctness point handled: an analyser is usually a LEAF TAP wired off a source and never connected to the speakers, so a naive render (which walks backwards from the destination) would never evaluate it - the context now tracks its analysers as extra render roots. The FFT is a scalar-f64 radix-2 transform in a fixed order, so it inherits the engine's determinism: the same graph yields byte-identical frequency data every run, proven as a gate invariant alongside a spectral-peak test (a 430.66 Hz tone peaks in exactly its bin). Media moves 35 to 36; overall holds at 51.