Fix Gemma 4 singleton expert prefill and prepare 0.2.4

[bsbodden]

What changed

• fall back to the Java projection kernel when Gemma 4 MoE batched prefill routes exactly one token to a quantized expert
• retain the native kernel for eligible batches of two or more
• add a Q4_K regression covering a batch-only injected kernel
• prepare the coordinated Models 0.2.4 patch release metadata

Root cause

Gemma 4 enables native batched prefill after validating representative two-token projections. MoE routing can still produce a one-token batch for an individual expert. The native kernel intentionally rejects singleton work unless native decode is enabled, but the Gemma decoder treated that eligibility decision as fatal instead of using its existing Java projection path.

User impact

The qualified ggml-org/gemma-4-26B-A4B-it-GGUF Q4_K_M artifact can now prefill prompts through ModelJars' default native backend without throwing loaded Gemma 4 batched-prefill kernel rejected Q4_K projection ... when an expert receives one routed token.

Validation

• ./gradlew :backend-java:test --tests com.integrallis.models.backend.purejava.gemma4.Gemma4ForwardPassTest
• ./gradlew spotlessCheck :backend-java:test
• ./gradlew verifyReleaseMetadata