Freeing memory from a finalizer waits for the GPU #717

Freeing memory from a finalizer waits for all GPU work in the process. The finalizer of an array that isn't sub-allocated (device memory over 1 MiB, and all shared and host memory) frees it with zeMemFreeExt and the BLOCKING_FREE policy (src/pool.jl). Our kernels declare indirect access, so NEO considers every allocation in use by every kernel, and the free waits until everything submitted so far has executed, not only the work using that buffer.

That stalls whatever thread happens to collect garbage, and deadlocks when the GPU work waits for a task on that thread. The test suite already works around it by disabling GC while a kernel waits on the host (test/execution.jl, "freeing memory may wait for the stream on some stacks").

Measured on an Iris Xe (NEO 26.35), with a ~1 s kernel running on another task's stream, and the release triggered from a finalizer on thread 1 (medians of 7):

With the stream waiting on an event that a task on thread 1 signals, collecting a host array (3/3), a 64 MiB shared array (3/3) or a 64 MiB device array (2/3) on that thread hangs until a watchdog signals the event.

Unlike with CUDA (JuliaGPU/CUDA.jl#3318), launches on other threads aren't held up, and only memory frees and the busy immediate-list drain wait. So rather than deferring every release, it should be enough to free with DEFER_FREE on the rolling stack. NEO keeps deferred allocations on a list that it only processes on a later free, so reclaiming memory needs to trigger one. The LTS stack reportedly doesn't honor the free policies (see the comment in src/pool.jl), so there the blocking free, and the unbounded synchronize_all_streams before it, would have to move out of the finalizer instead.