fix: Drain all buffered hops per PhaseVocoder.process() call

[bhj]

This one seems pretty straightforward. Related to #87

Model used: Anthropic Fable 5

PhaseVocoder.process() handled at most one analysis hop per call, unlike the WSOLA Stretch stage, which loops until its input holds less than one hop. Callers that feed fixed-size blocks and call process() once per block — the worklet processors feed 128-frame render quanta — starve whenever a block spans more than one hop: at hop 64 (fftSize 512 / overlap 8) output ran at roughly half speed with every block underrunning, and at hop 128 (512/4, 1024/8) down-shifts starved the same way once the transposer's expansion pushed a block's worth of input past one hop.

process() now drains every whole buffered hop per call, matching the Stretch contract. Measured at the previously broken configurations: unity-pitch correlation against the input recovers from ~0.04 to 1.0 with zero underruns, and down-shift underruns collapse from every block to the normal knife-edge residual.