Support ONNX DynamicQuantizeLinear

[jesco-absolut]

Summary

Adds DynamicQuantizeLinear support to candle-onnx.

This implements the ONNX v11 function semantics for F32 input to U8 output:

• computes the adjusted min/max range including zero
• returns scalar y_scale
• returns scalar y_zero_point
• quantizes output to U8
• uses ties-to-even rounding to match ONNX / NumPy rint

Fixes #3290.

Tests

Added test_dynamic_quantize_linear covering:

• mixed positive/negative input
• all-negative input
• all-positive input
• zero-range input
• ties-to-even rounding behavior

Validation run:

cargo test --manifest-path candle-onnx/Cargo.toml --test ops test_dynamic_quantize_linear
cargo test --manifest-path candle-onnx/Cargo.toml --test ops -- --skip test_gelu_operation
cargo test --manifest-path candle-onnx/Cargo.toml --lib
cargo check --manifest-path candle-onnx/Cargo.toml
cargo fmt --manifest-path candle-onnx/Cargo.toml -- --check
git --no-pager diff --check

Known unrelated local failures:

• full cargo test --manifest-path candle-onnx/Cargo.toml --test ops still fails only in existing test_gelu_operation exact-float mismatch: 0.8413447 vs 0.8413448
• cargo clippy --manifest-path candle-onnx/Cargo.toml --tests -- -D warnings is blocked by pre-existing generated-code / existing eval.rs warnings