Fix GDN recurrent cache sizing

[LxYuan0420]

This PR is:

• To fix under-reported GDN recurrent memory sizing by accounting for float32 recurrent state (introduced in [Qwen3.5] C++ Metal kernel for hybrid SDPA + GDN paged inference(Stage C) #226).
• To add a unit test covering the corrected linear cache byte calculation.

Notes:

• Scope is limited to hybrid linear attention accounting in linear_cache_bytes_per_slot().
• SDPA KV sizing is unchanged.