[libcu++] Make resolving a default memory pool legal under stream capture - #11360

…ture

`__get_default_memory_pool` applies the library's retention policy to the
driver's default pool by reading (and, the first time, writing) the
release-threshold attribute. `cuMemPoolGetAttribute`/`SetAttribute` are
"potentially unsafe" driver calls: while the calling thread is in a global or
thread-local stream capture they fail with CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED
and invalidate the capture. The default-pool lookup itself is a pure query and
is legal under capture; only the policy step is not.

Any lazy first touch of a default pool under an active capture therefore threw:
the first `device_default_memory_pool()` of a process, a pool ref constructed
while capturing, and every consumer that resolves the pool on its allocation
path (cudax's locality-domain data places since NVIDIA#11202, which broke the
sharded graph-capture tests).

Fix: run the policy step inside a `__relaxed_capture_scope`, an RAII scope
around `cuThreadExchangeStreamCaptureMode` that puts the calling thread in
relaxed capture mode and restores the previous mode on exit. Capture mode is a
per-thread property, so this is independent of any stream, has no observable
effect when the thread is not capturing, and costs two cheap driver calls
(~50 ns measured). The attribute write executes immediately rather than being
recorded, which is the intent for a process-global setting. Nothing that
should be captured is issued inside the scope.

Adds `__driver::__threadExchangeStreamCaptureMode`.

Test (libcudacxx/test/.../resources/default_pool_under_capture.cu): begins a
capture in global and in thread-local mode, resolves the device default pool
through the public once-only accessor (first touch of the process, on purpose)
and through the helper, allocates from the pool inside the capture, and checks
the capture ends successfully with at least one node and that the policy was
applied. A final section proves the thread's mode is restored: an unsafe pool
query right after the accessor is refused again. Driver entry points only, as
the ccclrt fixture requires. Fails before the fix with "operation not
permitted when stream is capturing(900): Failed to get attribute for a memory
pool"; passes after. memory_pools, locality_domain_memory_pool and
device_memory_resource tests unchanged.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

…lt-pool accessor

Co-Authored-By: Claude Fable 5.1 <[email protected]>

Review feedback on NVIDIA#11360.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

… host compiler

Since CUDA 13.x `CUmemLocation::id` lives inside an anonymous union (next to
the `localized` struct), so `CUmemLocation{type, id}` brace-elides a
subobject and clang 20 promotes -Wmissing-braces to an error under -Werror
in the CUDA 13.3 + llvm20 "Build libcudacxx" job. Default-initialize and
assign the fields instead, matching the style used in memory_pool_base.h.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

Co-authored-by: Jacob Faibussowitsch <[email protected]>

…ectness)

clang-tidy (misc-const-correctness, WarningsAsErrors) flagged the two
`cuda::stream stream{dev};` locals in default_pool_under_capture.cu: the
streams are only ever read through the const `get()` accessor and passed
to `allocate`/`deallocate` by value as a `stream_ref`, so they can be
`const`, matching how the sibling memory_resource tests declare them.

Co-Authored-By: Claude Fable 5.1 <[email protected]>