Async and base infrastructure for cross-process shared memory.

[benvanik]

This is the first half of the remote HAL transport stack, split out to land independently. Everything here is in iree/async/, iree/base/, and iree/testing/ — no net or HAL changes yet.

Shared memory primitives

Adds iree_shm_* for cross-platform shared memory create/open/close/map with handle passing support (memfd on Linux, shm_open on macOS, CreateFileMappingW on Windows). Includes memory sealing via iree_shm_seal() for making regions immutable after population — uses kernel-level F_SEAL_* on Linux, VirtualProtect defense-in-depth on Windows, unavailable on macOS.

On top of that, a lock-free SPSC queue (iree_spsc_queue_*) designed to operate on caller-provided memory so it can live in a shared memory region. Monotonically increasing 64-bit positions with acquire-release ordering, cache-line isolation between producer and consumer, skip markers for wrap-free reads. This is the data plane for the SHM carrier.

Cross-process notification

Extends the notification system to support shared-memory epochs and cross-process wake across all three backends. The core change is an epoch_ptr indirection so notifications can point at either the inline epoch (local, zero behavioral change) or a caller-provided epoch in shared memory.

Per-platform wake mechanisms:

• io_uring: shared futex (FUTEX mode) or caller-provided eventfd (EVENT mode)
• POSIX/Linux: shared futex + caller eventfd for poll loop wake
• POSIX/macOS: poll() on wake fd (no futex, no cross-process condvar)
• IOCP: RegisterWaitForSingleObject bridges caller Event to IOCP; WakeByAddress is per-process (virtual-address keyed), so cross-process wake uses the SetEvent path

Shared buffer pool

Extends iree_async_buffer_pool_t with create_shared/open_shared for cross-process zero-copy. The atomic freelist (64-bit CAS, position-independent indices) lives directly in shared memory so both processes can independently acquire and release buffers. Header magic is written last as a commit step so openers never see a valid header with uninitialized freelist state.

Proactor performance improvements

• IOCP event waits: Replace polling-based event wait with NtAssociateWaitCompletionPacket for zero-overhead kernel-level event-to-IOCP association.
• IOCP carrier freelist: Eliminates per-I/O heap allocation in steady state by recycling carrier structs through an atomic slist freelist. Pool grows on demand, drained at destroy.
• IOCP active_carriers: Doubly-linked list for O(1) carrier removal (was O(n) list walk on every event wait completion).
• Timer insertion fast-path: Tail comparison before list walk turns the common case (monotonically-increasing deadlines) from O(n) to O(1).
• Inline progress callbacks: Proactors can register progress callbacks that run inline during the event loop, enabling adaptive polling strategies without full wakeup overhead.

Bug fixes

• Send data lifetime in io_uring: Data referenced by send SQEs submitted from recv callbacks could be a use-after-return if the submitting function's stack frame unwound before the kernel read the data. Fixed by copying send data inline when immediate submission is possible. New data_lifetime_test in the socket CTS.
• Axis failure propagation: When an axis operation failed, the failure status was silently dropped instead of propagating to semaphores waiting on that axis value.

Testing infrastructure

Adds a coordinated multi-process test harness (iree/testing/). A single test binary re-executes itself in different roles, with the launcher orchestrating spawn order, readiness synchronization, and result collection. Supports Linux, macOS, and Windows. First consumer is the SHM carrier cross-process tests on the remote-hal branch.

Also adds iree_async_primitive_dup()/close() for cross-process handle transfer — the async primitive layer's equivalents of dup()/close(), DuplicateHandle/CloseHandle, mach_port_mod_refs.