Lift derived `head_dim` into a shared NormalizedConfig base class

Context

Follow-up from PR #382 review — see this comment by @zhenchaoni.

Several HuggingFace model configs do not expose head_dim natively (e.g. BartConfig, MarianConfig, and per the comment also blip and trocr). To make these models work with PastKeyValueInputGenerator (which now reads normalized_config.head_dim unconditionally), PR #382 introduced per-model NormalizedConfig subclasses with a computed head_dim:

• src/winml/modelkit/models/hf/bart.py — _BartDecoderNormalizedConfig.head_dim
• src/winml/modelkit/models/hf/marian.py — _MarianDecoderNormalizedConfig.head_dim

Both implementations are identical:

@property
def head_dim(self) -> int:
    return self.hidden_size // self.num_attention_heads

Proposal

Lift the derived head_dim property into a shared NormalizedConfig base class (e.g. _DerivedHeadDimNormalizedConfig in a common module under models/winml/ or models/hf/) so that BART, Marian, and the upcoming BLIP / TrOCR configs can simply inherit it, instead of redefining the same property in each subclass.

Acceptance criteria

• New shared base class with the derived head_dim property in a common location
• _BartDecoderNormalizedConfig and _MarianDecoderNormalizedConfig inherit from it and drop their local head_dim overrides
• BLIP and TrOCR (when added) reuse the same base class
• Existing BART / Marian export tests still pass

Start by reading src/winml/modelkit/models/hf/bart.py and src/winml/modelkit/models/hf/marian.py, then inspect the normalized-config modules under models/winml or models/hf for a suitable shared location. Extract the identical derived head_dim behavior into a shared base, update both decoder configs to inherit it, and run the existing BART and Marian export tests to confirm they still pass.