mtmd: use causal attn for gemma 4 audio (+ small breaking change to mtmd)

[ngxson]

Overview

Continue #21421

Fix #21820

Fix #21816

For gemma 4, the text model only use non-causal (aka bidirectional attention) for vision input

Breaking change

mtmd_decode_use_non_causal now requires passing a second param, the current chunk

The chunk is optional and can be nullptr (default: assuming the current chunk is image).

In the case of gemma 4:

• Vision chunk requires non-causal
• Audio chunk requires causal (same as text input)

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: no