Fix CuTeDSL editable install setup script

[VaggelisGian]

Fixes #3416

prep_editable_install.py downloaded only the pure-Python nvidia-cutlass-dsl wheel and globbed it for lib/*.so. Since the wheel split, that wheel carries no files at all, so the script copied nothing ("0 lib files, 0 Python files copied") and a following pip install -e . failed with ModuleNotFoundError: No module named 'cutlass._mlir'.

The script now downloads the wheels that actually hold the pieces:

• nvidia-cutlass-dsl-libs-base==<release> provides the cutlass._mlir python package (dialect bindings glue),
• nvidia-cutlass-dsl-libs-cu12==<release> (or -cu13 via a new --cu13 flag, matching setup.sh conventions) provides the toolkit-native _cutlass_ir binding and libcute_dsl_runtime.so,
• both are pinned to the release version parsed from the main wheel, whose pin itself is taken from this tree's requirements.txt.

It copies _mlir into the local cutlass/ package (base glue merged with the toolkit-native files), bridges this tree's own _mlir_helpers into cutlass/_mlir_helpers where its relative imports expect them, places runtime libraries under cuXX/lib mirroring an installed wheel so the DSL library discovery finds them, removes a stale flat lib/ left behind by older runs that would shadow the new layout, and fails loudly if no native binding was extracted instead of reporting success.

Adds python/CuTeDSL/.gitignore for the generated paths so running the script does not dirty git status.

Test Plan:

$ cd python/CuTeDSL && python prep_editable_install.py
INFO: Summary: 65 _mlir files, 9 _mlir_helpers files, 3 runtime library files (cu12/lib)
$ python prep_editable_install.py --cu13
INFO: Summary: 65 _mlir files, 9 _mlir_helpers files, 3 runtime library files (cu13/lib)
$ python -m py_compile prep_editable_install.py    # clean

# sha256 of every extracted file compared against the wheel members: 65+9+3 files, zero mismatches

$ docker run --rm -v <repo>:/src python:3.12-slim bash -c \
    "cd /src/python/CuTeDSL && python prep_editable_install.py && \
     pip install numpy cuda-python typing_extensions && pip install -e . && \
     python -c 'import cutlass; import cutlass.cute as cute; print(\"ok\")'"
import ok: /src/python/CuTeDSL/cutlass/__init__.py
cute ok

Before this change the same flow ends at ModuleNotFoundError: No module named 'cutlass._mlir'.

Note: numpy, cuda-python and typing_extensions are needed at import time but pyproject.toml declares no dependencies; that pre-existing gap is untouched here. The editable install remains Linux-only at runtime because the extracted natives are ELF shared objects.