feat(core.utils): add persistent program caches (sqlite + filestream)

Convert ``cuda.core.utils`` to a package and add persistent, on-disk
caches for compiled ``ObjectCode`` produced by ``Program.compile``:

* ``ProgramCacheResource`` — abstract Mapping-like base class with a
  context manager, ``get``/``close``, and a ``pickle``-safety warning.
* ``SQLiteProgramCache`` — single-file sqlite3 backend (WAL mode,
  autocommit) with LRU eviction against an optional size cap.
* ``FileStreamProgramCache`` — directory of atomically-written entries
  (staged tmp file + ``os.replace``) for concurrent multi-process use,
  with best-effort size enforcement by mtime. Windows ``PermissionError``
  from ``os.replace`` is treated as a non-fatal cache miss.
* ``make_program_cache_key`` — stable 32-byte blake2b key derived from
  code, code_type, ProgramOptions, target_type, name expressions, cuda
  core/driver/NVRTC versions, and NVVM-specific fields (``extra_sources``,
  ``use_libdevice``) that would otherwise collide.

``sqlite3`` is imported lazily so the rest of the package is usable on
interpreters built without libsqlite3.

Includes single-process CRUD, LRU/size-cap, corruption-handling, and
multiprocess stress tests, plus an end-to-end test that compiles a real
CUDA C++ kernel, stores it in each backend, reopens the cache, and
verifies ``get_kernel`` still works on the deserialised ``ObjectCode``.
Public API is documented in ``cuda_core/docs/source/api.rst``.

Author: cpcloud

cuda_core/cuda/core/utils.py

@@ -0,0 +1,14 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from cuda.core._memoryview import (
+    StridedMemoryView,
+    args_viewable_as_strided_memory,
+)
+from cuda.core.utils._program_cache import (
+    FileStreamProgramCache,
+    ProgramCacheResource,
+    SQLiteProgramCache,
+    make_program_cache_key,
+)