fix: fall back to source build when version string is not PEP 440

[Chessing234]

Bug

CachedWheelsCommand.run() calls get_wheel_url() without catching InvalidVersion. In vendor PyTorch containers (e.g. NGC images with internal build IDs), torch.__version__ or torch.version.cuda can be a non-PEP-440 string such as gpgpu.<build-id>. packaging.version.parse raises InvalidVersion on those strings, which propagates uncaught and aborts the install entirely.

Reported in #947.

Root cause

# setup.py, CachedWheelsCommand.run()
wheel_url, wheel_filename = get_wheel_url()   # InvalidVersion escapes here
...
except urllib.error.HTTPError:
    # only catches HTTP failures, never reached
    super().run()

get_wheel_url() calls parse(torch.__version__) and parse(torch.version.cuda). Both raise InvalidVersion for non-standard version strings. The except urllib.error.HTTPError block that falls back to source build is never reached.

Fix

Import InvalidVersion from packaging.version and wrap get_wheel_url() in a try/except at the call site. When caught, fall back to super().run() (source build), consistent with how urllib.error.HTTPError is already handled.

try:
    wheel_url, wheel_filename = get_wheel_url()
except InvalidVersion:
    print("Non-PEP-440 version string detected. Building from source...")
    super().run()
    return

Why this fix is correct

The existing except urllib.error.HTTPError path already expresses the intent: if the prebuilt wheel cannot be obtained for any reason, fall back to a source build. InvalidVersion is another such reason. The fix keeps the same fallback behavior and does not change what wheels are downloaded or how they are built.