Make simsimd an Optional Dependency with Automatic NumPy/SciPy Fallback for Vector Distance Computation (#249)

## Summary

This PR makes `simsimd` an optional dependency within the
`langchain-weaviate` integration by introducing a robust
NumPy/SciPy-based fallback implementation for cosine distance
computation.

Systems with older glibc versions or restricted environments cannot
install `simsimd`, which currently results in import errors and prevents
the package from functioning. This update ensures full compatibility
across environments while preserving the accelerated vector math path
where available.

## Motivation

`simsimd` provides excellent SIMD-optimized vector math, but it is not
universally installable—especially on older Linux distributions where
glibc upgrades are risky or not feasible.

This PR resolves that by:

- Making `simsimd` optional
- Falling back gracefully to NumPy/SciPy when unavailable
- Preserving feature parity with existing behavior
- Improving reliability across diverse production environments

## Changes Introduced

 1. Optional `simsimd` Import
- `simsimd` import is wrapped with a `try/except`.
- If unavailable, a warning is logged and NumPy/SciPy fallback is used.

 2. New Internal `cdist` Function
- Provides unified API (`cdist(X, Y, metric="cosine")`)
- Uses `simsimd.cdist` when available  
- Falls back to `scipy.spatial.distance.cdist` or a custom NumPy
implementation

3. Updated `cosine_similarity`
- Now delegates cosine distance computation to the new `cdist` function
- Behaviour preserved for both accelerated and fallback paths

 4. Test Coverage
Added tests validating:
- Fallback execution when `simsimd` is unavailable  
- Correct distance outputs using NumPy fallback  
- Error when unsupported metrics are passed  

All new tests pass successfully.

 5. Documentation Updates
README updated with a new section describing:
- Optional `simsimd` acceleration  
- Installation instructions (`pip install
"langchain-weaviate[simsimd]"`)
- Behavior when running without acceleration  

 6. pyproject.toml Updates
- `simsimd` moved to optional dependency extras  
- Ensures backward compatibility and avoids installation failures


## Backward Compatibility

This update is fully backward compatible:

- Users with `simsimd` installed experience **no change** in behavior.
- Users without `simsimd` can now install and use the package normally.
- No public APIs were modified.


## Additional Notes

- This PR improves the portability and reliability of the library in
production environments.
- The fallback path ensures correct numerical output and alignment with
existing interfaces.
- Maintainers can extend this pattern to other optional acceleration
modules in the future.


## Checklist

- [x] Optional import and fallback logic added  
- [x] New `cdist` function implemented  
- [x] Updated `cosine_similarity` to use unified distance API  
- [x] Tests added and passing  
- [x] README updated  
- [x] pyproject.toml updated  
- [x] Code linted and type-checked  

---

Thank you for reviewing this PR. I’m happy to incorporate any
suggestions or refinements.

---------

Co-authored-by: Duda Nogueira <dudanogueira@gmail.com>