fix(smolvlm): preserve processor-owned dimensions for no-resize packages - #2984

Problem

A real native SmolVLM training run completed and registered its model, but prediction failed with HTTP 500. The writer correctly recorded network_input.training_input_size: null because the dataset had no platform resize. The active inference-models reader rejected that value as an invalid TrainingInputSize before loading the model.

Change

Add a Smol-specific parser for the explicit processor-owned input contract. It accepts a null training size only with dynamic any-size RGB input and an identity platform transform, then uses the existing raw-image path into the packaged Hugging Face processor. It does not invent a resolution, alter trained artifacts, disable quantization, switch reader backends, or relax the shared fixed-size schema. Existing fixed-size packages retain their parser and preprocessing behavior. Nonidentity transforms with no size remain rejected rather than silently discarded.

Verification

• 12 focused tests passed against the exact installed ARM reader dependencies in a disposable CPU process.
• Tests cover native null-size metadata, unchanged fixed 640x480 parsing, malformed/missing size, static-input rejection, transformation rejection, and original variable image dimensions reaching the processor.
• The schema-only candidate preserves all 112 parent layers and runtime configuration, adding two Python source files.
• Its native same-model resume accepted the metadata, then exposed a second dependency defect: installed bitsandbytes 0.47.0 has no CUDA13 ARM library. That real failure is preserved separately.
• The isolated combined candidate adds the official bitsandbytes 0.48.1 ARM wheel, without changing Torch, model quantization, precision, backend, or other dependencies. Its build profile and hash-pinned requirements are included here.
• A real CUDA13 ARM NF4 quantize/dequantize probe passed on the canary MIG slice (relative error 0.09224). This is kernel validation, not full-model prediction validation.
• First combined candidate sha256:8ed0a5f2c597c69979910aff7a6242a8ceeb7be7d64acd24a33828317727c4ff failed containerd extraction because its replacement layer omitted explicit directories. That failure is preserved. Corrected candidate sha256:3fb47f5a20b59eb22dc613034e750ab665052274ac707645013914c3c062c836 passed actual Pod source/dependency/readiness checks.
• Supported native same-model suite resume PASSED: all 7 held-out images/prompts returned nonempty text in 70.09 seconds, preserving generation defaults. Training remained the original completed 14-epoch run, not a retrain. Both preceding real prediction failures are archived separately.
• Automatic evaluation is explicitly unsupported for text-image-pairs. This proves the native functional path, not matched hosted quality, numerical parity, or driver-hang prevention. Reader logs also warn that the installed processor's torchvision backend substitutes BICUBIC for LANCZOS; that existing dependency behavior remains a fidelity follow-up, not silently claimed equivalent.

Scope

Five-file, main-based source PR: reader helper, reader integration, regression tests, isolated build profile, and hash-pinned requirements. No shared inference deployment or benchmark training image is changed. These fix a writer/reader contract mismatch and an isolated CUDA13 dependency mismatch, not the NVIDIA driver hang. Native training 72e78311b43912927047 is reused; no retraining is needed to exercise these reader changes.