cuda.core: reject stream=None — require explicit stream everywhere

[leofang]

Summary

Several cuda.core APIs accept stream=None and silently fall back to default_stream() (or NULL). This makes the stream choice implicit and environment-dependent (CUDA_PYTHON_CUDA_PER_THREAD_DEFAULT_STREAM), which is error-prone. Users should always pass a stream explicitly — including device.default_stream when they want the default.

Design rule

stream should be a keyword-only argument with no default value (*, stream: Stream), so callers must always pass a stream explicitly.

APIs that already follow this convention:

• Buffer.copy_to(dst=None, *, stream: Stream | GraphBuilder)
• Buffer.copy_from(src, *, stream: Stream | GraphBuilder)
• Buffer.fill(value, *, stream: Stream | GraphBuilder)

APIs that need to change

stream=None → default_stream() (implicit fallback, remove it)

• MemoryPool.allocate() / deallocate()
• GraphMemoryResource.allocate() / deallocate()
• GraphicsResource.map()
• LegacyPinnedMemoryResource.allocate()
• _SynchronousMemoryResource.allocate()
• Device.allocate() (delegates to the above)

stream=None → NULL / legacy default (same issue)

• Kernel.max_potential_cluster_size(config, stream=None)
• Kernel.max_active_clusters(config, stream=None)

stream should become keyword-only

• GraphBuilder.launch(self, stream: Stream) — currently positional, should be (self, *, stream: Stream)

APIs that are fine (no change needed)

• launch(stream, config, kernel, *args) — stream is the 1st positional arg by design
• Buffer.close(stream=None) — None means "reuse the original stream", by design
• GraphicsResource.unmap(stream=None) / close(stream=None) — same, reuses the mapping stream

Proposed implementation

1. Have Stream_accept() raise TypeError when stream is None, so the check is centralized and cannot be forgotten.
2. Remove the = None default from every affected API signature.
3. Make stream keyword-only where it isn't already (except launch() where it's intentionally positional).

Prior art: CCCL's cccl-runtime

CCCL's modern CUDA runtime (libcudacxx, cuda:: namespace) follows the same explicit-stream principle:

The CUDA default (NULL) stream is not exposed as a first-class runtime object because it is tied to implicit per-device state and encourages hidden dependencies.

— docs/libcudacxx/runtime/cudart_interactions.rst

Concretely:

• cuda::launch(stream, config, kernel, args...) requires an explicit stream_ref as the first parameter, with no default.
• The cuda::mr::resource concept requires allocate(cuda::stream_ref, ...) / deallocate(cuda::stream_ref, ...) — no fallback to a default stream.
• cuda::stream_ref deletes constructors from int and nullptr, and deprecates the parameterless (default stream) constructor.

[leofang]

@Andy-Jost could you take this one too?