Modernize Pallas Splash attention kernel and define eager fallback routing in Tunix Gemma 4.

- Refactor Attention.block() into modular helper methods.
- Broaden Pallas Splash attention execution across rectangular sequences (kv_len > q_len) using memoized in-kernel computable masks (LocalMask / CausalMask).
- Define eager attention fallback routing (_eager_attention) with exact offset alignment for rectangular sequences.
- Modernize use_gqa property to properly support MQA (num_kv_heads=1).
- Add unit tests for mask predicate equivalence, block-size divisibility, and attention routing.

PiperOrigin-RevId: 963446160

Author: mariusdkm