Fix slowdown from utf conversion

[SimeonStoykovQC]

Motivation

4.5.8 (or 4.5.7) introduced a different utf16<->utf8 conversion (using wstring_convert instead of boost's conversion library), which considerably slows down the reading of string data - I observed 10x slowdown with 1m rows of characters.

Apart from the performance improvement, the changes are warranted because wstring_convert was deprecated in C++17.

Changes

• Migrate away from wstring_convert in favor of simdutf
  • similar performance to the old boost implementation (no formal benchmark in this PR - just local tests),
  • is easily buildable and conda-packaged: similar setup with ninja and cmake, see recipe
  • slightly more cumbersome to use - very open to any feedback and comments
• Include Windows CI

[xhochy]

@SimeonStoykovQC Can you rebase/merge with main?

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Comparison is base (10f7133) 98.81% compared to head (6e1e5a2) 98.86%.

Additional details and impacted files

```diff @@ Coverage Diff @@ ## master #400 +/- ## ========================================== + Coverage 98.81% 98.86% +0.05% ========================================== Files 94 10 -84 Lines 3030 352 -2678 ========================================== - Hits 2994 348 -2646 + Misses 36 4 -32 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[xhochy]

Please add a changelog entry.