- Updated the `ndarray_dot` and `ndarray_matmul` functions to utilize ndarray's optimized implementations for better performance.
- Improved error handling for shape and dtype mismatches in both operations.
- Refactored code to streamline view extraction and dimensionality conversion for Float32 and Float64 types.

- Optimize extract_view_as_* functions to avoid unnecessary mapv() for native types
- Consolidate comparison_helpers.rs and math_helpers.rs into op_helpers.rs
- Remove redundant shape returns from comparison and logical macros
- Update all comparison and logical operations to use new op_helpers module

… paths

Restructure all binary operation macros (arithmetic, bitwise, comparison, logical) into unified macros with zero-copy fast paths for same-type operations. When both operands share the same dtype, extract native views instead of converting, then broadcast and perform operations. All operations now properly broadcast.

- Reorganize modules: types/, macros/, helpers/, ffi/ with clear separation
- Move PadMode to types/pad_mode.rs and SortKind to types/sort_kind.rs
- Consolidate error handling: move panic utilities into ffi_guard module
- Consolidate indexing utilities: normalize_axis and normalize_index in helpers/indexing.rs
- Make FFI layer conditional on 'ffi' feature flag for selective compilation
- Update all internal imports (~50 files) to use new module paths
- Remove redundant helper files (stacking/helpers.rs, indexing/utils.rs)
- Rename validate_axis to normalize_axis for API consistency

Replace manual implementations of linspace, logspace, and geomspace with ndarray's Array::linspace, Array::logspace, and Array::geomspace methods.This leverages ndarray's optimized and well-tested implementations while simplifying our FFI code.

Refactor FFI array operations to use standardized view extraction helpers instead of wrapper methods. Removes redundant macros and simplifies the wrapper module by inlining operation logic directly in FFI functions.