Nothing "proven" but let me throw some things out there.

Since clock stretching does appear to work on Pi at the ACK bit, you could use a micro with at least one hardware I2C controller set to respond to the amplifier's address, and mimic every bus byte or condition on the other side as required, by bit-banging or a second I2C controller as available. Use clock stretching to hold off each side between bytes. This is straightorward and easy to work with, but slow. You can speed it up (and complicate it) if you cared to parse the read/write protocol of the amplifier chip, for example, to handle book/page changes asynchronously, or read-ahead from the amplifier when safe to do so, or cache written values...
If the Pi won't accept another master on the bus, and hold itself off during a transaction, you'll have to hold it off. You can do this from inside by giving the on/off/startup micro an output that tells the Pi to not initiate transactions. Even if the Pi is in master mode, its bus outputs in the idle state "should" be in undriven open-drain mode and shouldn't disturb whatever else might be happening on the bus. You'll need to handle the hold-off signal in the application or, if kernel drivers are issuing the commands, hack the i2c or ALSA(?) device driver.
Or, if your Pi side can handle retries when the amplifier disappears and reappears from the bus, use any I2C hot-plug transceiver to simply detach the Pi whenever the on/off/startup micro does its business. For best results you would need the micro to detect when the bus is in use by the Pi and only detach or attach when idle.