fix(onnx): reduce extrema over all dimensions for empty axes - #4024

Fixes #4023

ReduceMax and ReduceMin currently return the input unchanged for explicitly empty axes, even with the default noop_with_empty_axes=0. For example, an opset 18 model reducing [[3, -2, 7], [4, 6, 1]] with empty axes and keepdims=0 returns a [2, 3] tensor instead of scalar 7/-2.

Treat empty axes like omitted axes before choosing the existing reduction branch. This preserves modern noop_with_empty_axes=1, keepdims, nonempty axes, duplicate-axis handling and the existing zero-size-input error. The change is limited to the two extrema branches.

Add four integration tests exercising 192 serialized ONNX model combinations through simple_eval: opset 13 axes attributes; opset 18 optional axes inputs; empty/omitted axes; positive/negative and unsorted multi-axis controls; default/0/1 keepdims and modern default/0/1 no-op. Legacy graphs do not use modern parameters. Tests check shape, values and dtype.

Validation on macOS arm64, Rust 1.94.1, protoc 33.6:

• Unmodified evaluator: 18/192 matrix cases fail; patched evaluator: 192/192 pass.
• Independent ONNX checker + ReferenceEvaluator: 192 models pass; patched Candle matches all 192 serialized fixture outputs.
• Existing ONNX operation tests: 67 pass, one pre-existing GELU precision failure. The same failure occurs on unmodified main at tests/ops.rs:1919, comparing 0.8413447 and 0.8413448. A supplemental run excluding that one test passes all 67 existing tests and all four new tests, plus doc-test targets.
• cargo test --locked -p candle-core --lib --tests: 204 pass, one ignored.
• ONNX and core cargo build ... --all-targets: pass. ONNX and workspace format checks: pass.
• Strict ONNX Clippy fails with the same 10 diagnostics on the baseline and patched evaluator: six generated-protobuf documentation warnings and four existing evaluator warnings (comparison_to_empty, needless_borrow, clone_on_copy, wildcard_in_or_patterns). The new regression target passes Clippy when only those five already-observed lint categories are allowed. This supplemental check is not a strict-Clippy pass.

Core strict Clippy (cargo clippy --locked -p candle-core --all-targets -- -D warnings) also passes. Local CPU verification does not establish remote CI or other-platform results.

AI-assisted implementation and testing; final code and baseline exceptions independently reviewed before submission.