AUv3: lay the subview out at the agreed size, and serialize audio unit creation

[defiantnerd]

Closes #553. The issue holds two unrelated defects in one file; they are two commits here.

1. The subview ignored adjust_size. After the container consulted adjust_size (added in #510), the layout pass still framed the subview from the raw container bounds — so an iPad rotation to landscape gave 1024x600 → adjust_size → 800x600 → plugin lays out at 800x600 → frame forced back to 1024x600, stretched. setGUISize:height: now writes the adjusted size back into the cached _guiWidth/_guiHeight (that cache was previously only updated by gui_request_resize, so it went stale after every host-driven resize), and the layout pass reads the agreed size back rather than using bounds. Side benefit: a non-resizable plugin, where setGUISize: is never called, also stops being stretched to the container.

2. Two AudioUnits could be created. _bootstrapAttempted guards re-entry on the main thread only — not against the host's factory call, which arrives on an XPC worker. Both could see audioUnit == nil and build one; the property overwrite then released the bootstrap AU whose GUI already existed, so dealloc reached terminate() without gui->destroy() (a CLAP contract violation, and a JUCE assert), while _tryCreateGUI ran against the survivor whose _guiCreated was false — a permanently empty editor. A new -createAudioUnitOnce:error: on the base class holds one mutex, re-checks, and runs the caller's build block; both generated-code producers now funnel through it instead of doing their own unlocked check-then-act. No second AU is built, so there is no loser to tear down.

Trade-off worth a reviewer's attention. Taking a real lock newly exposes one hazard: a plugin doing dispatch_sync to the main queue from inside clap_plugin->init() would deadlock against a main thread parked on that lock. That is a plugin bug (init is [main-thread]), but it was not reachable before. The alternative — try_lock with a placeholder-size fallback — would reintroduce the wrong-initial-size bug #510 fixed. There is no dispatch_sync anywhere in auv3_audiounit.mm, so nothing in the wrapper itself can deadlock here.

Scope. auv3_audiounit.h gains a five-line declaration so the generated subclasses can see the new method; the alternative was duplicating it as a category in both producers with no compiler cross-check.

Verification. Compile-verified under the Xcode generator: clap-first-distortion_auv3 and clap-first-distortion_auv3_standalone both build after each commit, and the regenerated generated_auv3_entrypoints.hxx in the build tree was confirmed to contain the new body, so the build-helper path is genuinely exercised. The iOS .hxx.in template is textually identical but was not compiled — no iOS toolchain here. Neither fix is runtime-tested: both need a physical iPad in AUM or GarageBand, and the race in commit 2 needs a slow-loading plugin to have any chance of reproducing.