Fix CWD-relative CUTLASS C++ compilation cache (#3404)

[notkisk]

Description

Fixes #3404.

cutlass_cppgen previously stored compiled artifacts in compiled_cache.db relative to the process working directory. This allowed an attacker-
controlled database to supply a native library that was loaded through ctypes.CDLL.

This change:

• Stores the cache under CUTLASS_CPPGEN_CACHE, XDG_CACHE_HOME, or ~/.cache/cutlass_cppgen.
• Uses restrictive permissions for the cache directory and database where supported.
• Stops forcing temporary CUDA and shared-library files into the working directory.
• Cleans up temporary files and SQLite connections deterministically.
• Adds regression tests for cache placement, permissions, and temporary library handling.

Existing CWD-local caches are intentionally no longer used and will be rebuilt in the new per-user cache location.

Testing

CUTLASS_CPPGEN_CACHE=/tmp/cutlass_cppgen_test_cache \
PYTHONPATH=/tmp/cutlass_cppgen_test_deps:python \
python -W error::ResourceWarning -m unittest test/python/cutlass_cppgen/test_cache.py

Result: Ran 4 tests ... OK

[notkisk]

@hwu36 @jackkosaian