Foreword

This PR makes substantial changes to the current implementation's structure, but it is byte-accurate with the original and the C implementation. Most changes are in the structure and control flow rather than the actual floating point operations. Reviewing the diff on GitHub itself my be challenging, so I recommend checking out the before and after in their entirety.

Objective

• Simplify a somewhat non-standard Rust crate into something more idiomatic.
• Remove libm as a dependency, allowing this crate to stabilize.

Solution

• Removed several custom sorting implementations, opting for core::slice::sort. Since the original implementations were correct, we can use the language provided sorts instead.
• Removed the custom hashtable implementation, opting to instead use alloc::collections::BTreeMap.
• Removed libm as a dependency and removed all features, making this crate purely no_std. Instead, users must provide a VectorSpace implementation alongside the implementation of Geometry. This has the two-fold benefit of simplifying internal code and allowing users the freedom to choose their maths library (avoiding issues where glam was pegged to a particular version for example).
• Reorganised code to favour slices over manually tracked buffers/offsets/lengths.
• Used iterator constructs to provide clearer structure around intent.
• Added documentation where only // comments existed.
• Used lifetimes and core::slice::split_at(_mut) to track the relationship between indices and buffer regions.

Notes

This PR removes about a third of the code from this crate while adding customization options for the end user and paving a clear path to stabilization. But, it is also a substantial departure from the structure of the original C implementation. As such, I expect this to be somewhat controversial.