feat(server meta): expose SWA geometry in /v1/models meta

[marksverdhei]

Closes #123.

What

Adds 3 new fields to the /v1/models meta object so clients (heierchat VRAM panel) can compute exact KV bytes/token for SWA-hybrid models (gemma-4 12B/31B, gemma-4-12b-256k@gem) instead of the current safe-upper-bound estimate.

New meta fields (per loaded model)

Field	Source	Example (gemma-4 12B)
n_swa	llama_model_n_swa(model)	1024
n_swa_layers	new llama_model_n_swa_layers(model)	40
swa_type	new llama_model_swa_type_name(model)	"standard"

New public API

LLAMA_API int32_t    llama_model_n_swa_layers(const struct llama_model * model);
LLAMA_API const char * llama_model_swa_type_name(const struct llama_model * model);

Exact KV formula the FE will compute

kv_per_tok(ctx) = (k_gqa + v_gqa) * kv_type_bytes * (
                    n_full_layers
                  + n_swa_layers * min(ctx, n_swa) / ctx )

where n_full_layers = n_layer - n_swa_layers.

Verification

• CPU-only build (-DGGML_CUDA=OFF): compiles clean, zero warnings
• Purely additive — no behavior change until consumers read the fields

Files changed (4, +36 lines)

• include/llama.h — 2 new API declarations
• src/llama-model.cpp — implementations (layer-count loop + enum→string)
• tools/server/server-context.h — 3 new fields in server_context_meta
• tools/server/server-context.cpp — populate + emit in /v1/models JSON