add TransformerConfig with typed attention and MLP variants

[gerardcode]

feat: add TransformerConfig with typed attention and MLP variants

Introduce TransformerConfig struct to ModelConfig with typed constants
for attention type (mha, gqa, mla) and MLP type (dense, moe), along
with key hyperparameter fields (numLayers, numAttentionHeads, numKVHeads,
hiddenSize, intermediateSize). Update config-schema.json and docs/config.md
accordingly.

This lays the groundwork for the unified Transformer specification,
enabling inference engines to auto-detect model variants without
per-model adaptation.

[gemini-code-assist]

Code Review

This pull request introduces a transformerConfig object to the model configuration across documentation, JSON schema, and Go specifications, enabling the definition of architectural parameters like attention and MLP types. Feedback recommends clarifying default values for attentionType and mlpType and resolving a logical inconsistency regarding numKVHeads in GQA configurations. Additionally, the reviewer noted that the schema lacks specific parameters required for MoE and MLA variants, such as expert counts and latent ranks.

[gemini-code-assist]

It is recommended to explicitly state the default value for attentionType when it is omitted from the configuration. Typically, this defaults to "mha".

Suggested change

	The attention mechanism variant. Supported values:
	The attention mechanism variant. Defaults to "mha". Supported values:

[gemini-code-assist]

It is recommended to explicitly state the default value for mlpType when it is omitted. Typically, this defaults to "dense".

Suggested change

	The feed-forward / MLP layer variant. Supported values:
	The feed-forward / MLP layer variant. Defaults to "dense". Supported values:

[gemini-code-assist]

The description for numKVHeads states that omitting it implies standard MHA. This creates a logical conflict if attentionType is explicitly set to "gqa". The documentation should clarify that this field is required for GQA configurations and should be less than numAttentionHeads.

[gemini-code-assist]

The comment for NumKVHeads suggests that omitting it always implies MHA, which contradicts the AttentionType field if it is set to GroupedQueryAttention. The documentation should reflect that this field is essential for GQA.

[gemini-code-assist]

The TransformerConfig definition includes "moe" and "mla" variants but lacks the parameters necessary for an inference engine to actually configure them (e.g., numExperts and numExpertsPerToken for MoE, or latent ranks for MLA). To support the goal of auto-detection mentioned in the PR description, these fields should be added to the schema.