{CUDA Decoding} Restore SW decoder fallback in xprsDecoderMaker

[YLouWashU]

Summary:
A wheel built with ENABLE_NVCODEC=ON segfaults on any machine without an NVIDIA driver. The GitHub CI "Test wheel" job crashes with exit 139 after:

Cannot load libcuda.so.1
[XPRS][WARNING]: Loading CUDA functions failed
[DecoderFactory][WARNING]: Could not create a decoder for 'H.265'!

With no driver, getNvCodecContext() throws while probing the HW entries and enumDecoders() catches it, returning a non-OK result. The SW hevc/h264 decoders were already collected before the throw (they precede the HW entries in kPreferredDecoderImplementations), so the list is usable. xprsDecoderMaker bailed on the non-OK result and discarded that list, so no decoder was created and the null decoder was dereferenced downstream. Note the HW decoders never enter the list at all: the throw happens before codecs.push_back.

Three changes:

• Ignore the enumDecoders() result and fail only when the decoder list is genuinely empty. This is what fixes the crash.
• On decoder init() failure, continue to the next candidate instead of returning, so the HW->SW fallback actually runs. This matters on a machine that does have a GPU, where enumeration succeeds but init() fails (NVDEC session limit, VRAM pressure).
• Return nullptr rather than decoder when the loop is exhausted. Raised by DevMate: with the continue above, a candidate that initialized but failed decode() could otherwise be returned by a later iteration's fall-through. A non-null return is also load-bearing in DecoderFactory::makeDecoder, which stops at the first maker that returns anything, so handing back a known-bad decoder would prevent other registered makers from being tried. This also removes a pre-existing path to the same stale return.

The GPU path is unchanged.

Reviewed By: kongchen1992

Differential Revision: D114806413

[meta-codesync]

@YLouWashU has exported this pull request. If you are a Meta employee, you can view the originating Diff in D114806413.