refactor(core.utils): raw-binary filestream cache, drop SQLite + pickle

* Delete SQLiteProgramCache entirely; FileStreamProgramCache is the only
  backend. Drops the sqlite3 import and ~500 lines of WAL/VACUUM/eviction
  plumbing.
* Cache stores raw binary verbatim (no pickle, no JSON, no record header)
  so entry files are directly consumable by external NVIDIA tools
  (cuobjdump, nvdisasm, cuda-gdb, ...).
* API is bytes-in / bytes-out: cache[key] = value accepts bytes-like or
  ObjectCode (path-backed too -- the file is read at write time so the
  cache always holds the binary content, not a path); cache[key] returns
  bytes. Callers reconstruct ObjectCode themselves with the code_type
  they already know. symbol_mapping is intentionally not preserved.
* Eviction is now true LRU by st_atime: every successful read calls
  os.utime(path, ns=(now, mtime_ns)) to bump atime regardless of mount
  options or NTFS LastAccess defaults. Replaces the previous st_mtime
  sort (effectively FIFO).
* Default cache directory follows OS conventions when path is omitted:
  $XDG_CACHE_HOME/cuda-python/program-cache (Linux),
  ~/Library/Caches/cuda-python/program-cache (macOS), or
  %LOCALAPPDATA%\cuda-python\program-cache (Windows).
* Bump _FILESTREAM_BACKEND_SCHEMA 2 -> 3 so existing pickle-format caches
  are wiped on first open.
* Tests rewritten: drop SQLite suite, drop pickle/corruption tests,
  add atime-LRU regression, add path-backed-accept regression, add
  raw-binary on-disk regression, add multiprocess eviction-race test
  (rewriter vs. churner) for the stat-guarded eviction path.

Author: cpcloud

cuda_core/cuda/core/utils/__init__.py

@@ -10,20 +10,18 @@
 __all__ = [
     "FileStreamProgramCache",
     "ProgramCacheResource",
-    "SQLiteProgramCache",
     "StridedMemoryView",
     "args_viewable_as_strided_memory",
     "make_program_cache_key",
 ]
 
 # Lazily expose the program-cache APIs so ``from cuda.core.utils import
-# StridedMemoryView`` stays lightweight -- the cache backends pull in driver,
+# StridedMemoryView`` stays lightweight -- the cache backend pulls in driver,
 # NVRTC, and module-load machinery that memoryview-only consumers do not need.
 _LAZY_CACHE_ATTRS = frozenset(
     {
         "FileStreamProgramCache",
         "ProgramCacheResource",
-        "SQLiteProgramCache",
         "make_program_cache_key",
     }
 )