fix(core): Match the transposer rate to the stretch stage's realized tempo

[bhj]

NOTE: This will conflict with the changes in #84 but I can rebase if that gets merged first. Thanks.

Model used: Anthropic Opus 5

SoundTouch asks the Stretch stage for 1 / pitch and, independently, sets the Transposer to pitch, assuming the stage delivers the tempo it was given. The built-in Stretch does, via its fractional skipFract accumulator. PhaseVocoder does not: its synthesis hop is round(Ha / tempo) whole frames, so it realizes Ha / round(Ha / tempo) instead — off by up to 1 / (2 * Ha), or 0.11% at the default 2048/4.

The stages then disagree about duration, and the error accumulates for as long as the pipeline runs: roughly 270ms of slip over a four-minute track. A real-time consumer feels it either as latency creeping upward or as the output buffer draining until it hands back silence, depending on which way the hop rounded.

Two halves:

• PhaseVocoder snaps tempo to a hop it can actually deliver and reports the snapped value, so the stage no longer claims a tempo it will not honour. Costs the same fraction of pitch accuracy it used to lose to drift, about 0.02 semitones, which is inaudible.
• SoundTouch derives _rate from the tempo the stage reports rather than from virtualPitch, via the existing public StretchPipe.tempo getter. No interface change. The built-in Stretch returns what it was given, so _rate stays exactly virtualPitch there and that path is untouched.

Two PhaseVocoder specs asserted the old exact tempo round-trip and now assert the snapped value.