fix(python): avoid CPU bf16 compile flake

[ddupg]

Summary

• Fix the Windows Python CI flake in test_l2_distance_f16_bf16_cpu.
• Failure reference: https://github.com/lance-format/lance/actions/runs/28589246239/job/84768429390
• pairwise_l2 already upcasts CPU float16 before matmul; extend that same path to CPU bfloat16.
• The result is still cast back to the original dtype, while the CPU matmul/accumulation runs in float32.

Tests

• git diff --check
• Not run: env UV_CACHE_DIR=/private/tmp/uv-cache-lance-ddu259 make install PYTHON=3.12 failed during the local pylance native build with Operation not permitted while writing a cargo .rlib under python/target.