[CUDA] Use native events for GPU fence waits

[strayberry]

#4400
Use native CUDA events for GPU-to-GPU dependencies to avoid host polling, while retaining counted AtomicEvent synchronization for CPU dependencies.

Fence simultaneously maintains:

• producer stream;
• MLX Event with value fixed to 1;
• AtomicEvent maintaining an incrementing count.

Synchronization rules:

Producer	Consumer	Implementation
GPU	GPU	producer Event::set_value(1), consumer stream asynchronous wait
GPU	CPU	counted AtomicEvent
CPU	GPU	counted AtomicEvent
CPU	CPU	counted AtomicEvent

Fixing the value to 1 makes CUDA Event select native cudaEvent_t, avoiding falling back to counting device-memory events. The consumer GPU stream establishes order via cudaStreamWaitEvent without blocking host.

Constraints

• CPU side cannot be changed to a value-one event, otherwise it breaks repeated update/wait semantics;

• GPU→GPU must use the passed-in consumer stream;

• Cross-device synchronization or dependencies involving CPU continue using AtomicEvent;

• Modifications are strictly limited to MLX CUDA provider, without altering Metal, MCI format, or model weights;

• ☑️ I understand it is strictly prohibited to use AI to write PR description

• AI usage disclosure:
  analysis and review code