MR: tinyusb_net — sync TX backpressure + thread-safe send_sync

[tore-espressif]

Why

• NCM busy: do_send_sync used to fail the frame when tud_network_can_xmit was false, returning ESP_FAIL. LwIP can react badly (e.g. TCP stalls). We must not block/spin on the USB task either.
• Concurrency: One global packet_to_send + stack packet + old handshake was not safe if several tasks called tinyusb_net_send_sync.
• Timeouts: Needed one clear deadline for “wait for TX” + “wait for deferred xmit done”.

What changed

• Caller runs tinyusb_net_wait_tx_ready(): poll tud_network_can_xmit(len) under vTaskSetTimeOutState / xTaskCheckForTimeOut, yield + periodic 1 ms delay; on deadline → ESP_ERR_TIMEOUT (no defer).
• USBD task (do_send_sync): tud_network_xmit only, after assert(tud_network_can_xmit) (debug). No FIFO-full ESP_FAIL from this path.
• Sync path: sync_mutex + stack sync_tx_job_t + shared sync_done_sem completion; drop event group / old buffer_sema / global packet_to_send. Drain done-sem before defer to avoid a stale Give. Mutex and completion Take use ticks_remaining so one budget covers the whole call.
• tinyusb_net.h: notes for full-call timeout and ESP_ERR_TIMEOUT. Async behaviour unchanged in design.

TODO

• Test with esp-idf example tusb_ncm
• Test with esp-idf example sta2eth

Related

Closes espressif/esp-idf#18374