Fix O4d CPU container fallback transfers and ILE transfer separators

[oshaughnessy-junior]

CPU-only ILE and CALPILOT jobs using a container family in container-universe mode could hold because their transfer lists still referenced GPU capability attributes absent from CPU slots. Select and transfer the fixed fallback image for these jobs, while preserving ILE --gpu --force-xpy --vectorized arguments for its CPU NoLoop path. GPU jobs continue to transfer the selected full image URL and pin the effective transfer list.

Also separate string-valued ILE transfer inputs from added container files with commas; previously the first added input could be concatenated onto the caller's filename. Existing O4d calibration-reweighting fallback handling remains in place.

Validation: container writer regressions and actual HTCondor effective-job-ad checks cover GPU/CPU ILE (list and string inputs) and CALPILOT. All 34 tests passed in the production environment, including all six real condor_submit -dry-run cases (none skipped). Candidate source hashes were verified against the published commit. GitHub CI is unavailable.

This follows the O4c correction for O4d. Independent adversarial review passed, including multi-file and empty-input transfer cases and precedence when both container selection modes are enabled. A bounded live CPU container probe using the identical writer source completed successfully with NoLoop arguments retained; it exercised scheduling, fallback image transfer and container startup with a small test image, rather than scientific inference. GPU capability upper-bound policy and the optional runtime selector's assigned-device behavior remain separate known limitations.

[interhost-cross-review-canary-v2]

Independent automated review completed at the recorded exact commit. Detailed findings were withheld from public output by the private-context egress policy and require private human declassification.