Investigate EmbeddingGemma fp16 CPU regression on Linux ARM64

[Copilot]

Description

• Scope: Investigation only; no source or test changes.
• Finding: 1.24.1 introduced Linux ARM64-native fp16 arithmetic, excluded on Apple ARM64. These registrations can replace float32 fallback regions and change intermediate rounding even with ORT_DISABLE_ALL.
• Unresolved: No incorrect kernel arithmetic demonstrated. First-divergence traces and saved execution graphs are needed before proposing a fix.

Motivation and Context

EmbeddingGemma-300M’s fp16 export produces NaN embeddings on Linux ARM64 starting in 1.24.1, while the same batch succeeds on 1.23.2 and macOS ARM64. Reported residual overflow near fp16’s range limit makes changed precision boundaries a plausible explanation.

• Fixes EmbeddingGemma-300M fp16 CPU inference regression: NaN output on Linux aarch64 since 1.24.1 #32600

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.