Remove build-host CPU variants

[jeffhuen]

Summary

• remove Linux, macOS, and Windows build-host CPU probes from RustlerPrecompiled loading
• ship 30 portable NIF artifacts instead of 30 baseline plus 15 x86-64-v3 variants
• remove the duplicated 32-byte scanner and encoder loops instead of adding runtime unsafe dispatch
• update release instructions, architecture notes, and the 0.4.1 changelog

Performance decision

A four-leg baseline / AVX2 / x86-64-v3 / baseline comparison ran through Criterion and the production decode and encode APIs on AMD Zen 3 and Zen 4 hosts. AVX2 reduced isolated scanner time by about 35-41%, but decode and encode showed no consistent end-to-end gain outside BEAM benchmark drift. The result does not justify carrying runtime dispatch and an unsafe target-feature boundary.

Benchmark evidence: https://github.com/jeffhuen/RustyCSV/actions/runs/31866641117

Verification

• Rust fmt, default/all-feature clippy, and 125 Rust tests
• 54 core plus 4 encoder Miri tests
• 494 ExUnit tests plus 5 properties
• Credo strict and Dialyzer
• docs with warnings as errors
• Hex retirement and RustSec audits
• unpacked Hex package source build
• normal precompiled baseline load and forced source build