k_timer: remaining semantic divergences from upstream (remaining_get type, status_sync busy-poll)

[swoisz]

Tracked divergences on the k_timer API, post-#17:

• k_timer_remaining_get() return type: Boreas int64_t, upstream uint32_t ms. Consumer code doing uint32_t r = k_timer_remaining_get(...) silently truncates today. Widening is the safer divergence but should either match upstream or carry an explicit @note.
• k_timer_status_sync() busy-polls: a sem-based implementation was attempted and reverted — the embedded stack-local struct k_sem blocking on K_FOREVER corrupts FreeRTOS scheduler state on ESP32-S3 (see the k_sem corruption issue). Doxygen documents the busy-poll. Re-attempt once the k_sem issue is fixed.
• Callback context visibility: with CONFIG_K_TIMER_DISPATCH_ISR=y (default) expiry runs in ISR context matching upstream; the task-context fallback path deserves a louder callout in the zkernel README for consumers porting Zephyr code.

[swoisz]

Related: #21 (the k_sem bug blocking the sem-based status_sync re-attempt)

[swoisz]

Now doubly unblocked: the k_sem corruption that forced the busy-poll was root-caused and fixed (#21/#18), and k_sem itself is now the notification-backed implementation (#41) whose blocking path the sem-based status_sync design originally wanted. Natural pairing: land this alongside the stress-hardening work in #43.