Join the second-core render with a dedicated semaphore

[rt-rtos]

esp_render_on_cores() joins the second-core worker with a plain task
notification on the calling task's default notification index. That counter is
not private to the join: the calling task belongs to the app (with
multithread = 0 it is whatever task runs amy_update()), and pacing such a
task with task notifications is idiomatic FreeRTOS - a timer ISR giving one
notification per audio block, for example. When a tick lands inside the join
window, the join returns on the tick instead of the worker: the caller runs
amy_fill_buffer() while the worker is still writing fbl[1] and combines a
half-written buffer. The worker's real "done" give then miscounts the app's
next wait, desyncing its clock. It presents as intermittent, load-dependent
garbling that only appears in multicore configs.

This replaces the caller-side join with a dedicated binary semaphore, which
nothing else can signal, and retires amy_render_task_done_handle. The worker
kick keeps its notification: the worker task is AMY-internal and nothing else
notifies it.

Indexed notifications (xTaskNotifyGiveIndexed on a non-zero index) would
also work but require configTASK_NOTIFICATION_ARRAY_ENTRIES > 1, which the
supported cores don't all guarantee; the semaphore is portable.

Verification

A/B on ESP32-P4 (ESP-IDF 6.1, multicore=1, multithread=0), with the
amy_update() task paced by a GPTimer ISR giving one default-index task
notification per block - the pacing shape described above. To make the
overlap deterministic, the second-core render window was stretched past the
5.33 ms block period so every join spans a tick; a worker-busy flag checked
at join return counts early releases.

• Unpatched join: 23,929 early releases in a 2-minute run (~180/s,
  essentially every block), the first within 3 s of boot, with the expected
  knock-on desync as the join consumes the app's pacing ticks.
• With this patch: zero early releases; a separate 17-minute soak
  (194,000 joins) ran with no deadlock and no missed completion.

Originally observed on ESP32-S3 with the same GPTimer pacing: intermittent,
load-dependent dual-core corruption that stopped once the join no longer
shared the task's notification counter. Compile-verified on ESP32-S3
(ESP-IDF 6.0.2) as well.

[bwhitman]

⛓️ tulipcc integration PR opened

This merge was pinned into tulipcc for full-system CI: shorepine/tulipcc#1337

Test it there and merge that PR to move tulipcc onto this AMY.