This fixes warnings that are emitted when building with the `-analyze`
flag enabled.

When normalisation is disabled, the Convolution will now adjust the gain
of the IR using the ratio of the source and destination sampling rates.
This should keep the output level constant when the Convolution's
sampling rate is changed.

Previously, if `loadImpulseResponse` was called before `prepareToPlay`,
the IR wasn't guaranteed to have loaded before the first call to
`processSamples`.

Now, we flush the queue of pending IR-load commands during
`prepareToPlay`, which should ensure that the most recently-loaded IR is
ready to use immediately.

The startThread/stopThread calls were moved to prevent a thread
sanitizer warning about a race on the vtable of Impl.