gemma4/gguf: accept a scalar attention.head_count_kv

[Rhycecomley]

Fixes #63.

parse_gguf_config treats gemma4.attention.head_count_kv as a per-layer array and indexes
it directly. llama.cpp collapses that array to a single value when every block shares it,
which the official google/gemma-4-E2B-it-qat-q4_0-gguf
checkpoint hits (head_count_kv == 1 across all 35 blocks), so ft serve dies during
_adjust_config with:

TypeError: 'int' object is not subscriptable

This broadcasts a scalar to one entry per block before the SWA/full lookups. Arrays are
unaffected — the isinstance guard only fires on the scalar form.

Verification

Before: ft serve --model gemma-4-E2B_q4_0-it.gguf raises the TypeError above and the
frontend shuts the API server down.

After: config parsing completes and startup proceeds past this point.

Note my machine can't finish a full generation run on this checkpoint for unrelated reasons
(4 GB VRAM), so I've verified the parse fix rather than end-to-end inference — worth a second
look from someone with a supported card.

[Rhycecomley]

Scope note: as detailed in #63 (comment), this makes a scalar head_count_kv parse, but it does not make the E2B checkpoint loadable — the next line hits a per-layer feed_forward_length that ModelConfig.intermediate_size can't represent. Please treat this purely as a robustness fix for valid-but-uniform GGUF metadata, not as E-series support.

[mkornreich]

Does this relate to my PR? See #59