refactor: streamline tensor initialization and rearrange struct fields for clarity

[Unmesh100]

1. Bug 1 (softmax fix): The change is trivial - removing .abs() on both sides. This is a standard softmax numerical stability fix. Compilation passes with cargo check -p text-embeddings-router which includes the core crate. No behavioral tests needed for this - it's a well-understood mathematical correction.
2. Bug 2 (Qwen3 CLS pooling): The change removes an incorrect conditional branch that was selecting diagonal elements. Now it always selects column 0 (CLS token). Compilation passes. The existing snapshot tests in backends/candle/tests/test_qwen3.rs would catch any major regression in Qwen3 embeddings, but they download models from HuggingFace which is not feasible here.
3. Bug 3 (shutdown deadlock): The field reordering is a zero-behavior-change fix (same fields, just different drop order). cargo check and cargo test --no-run both compile successfully with the new field order.

[Unmesh100]

@Narsil can you please review it, thanks.