`batch_cast` does not use SIMD intrinsic when available on Intel

[AntoinePrv]

Plenty of conversion are available on Intel.
For instance

• Upcast integers: _mm_cvtepi32_epi64 on SSE i32 -> i64
• Convert ints to floats: _mm_cvtepi32_ps on SEE i32 -> f32
• Downcasting has more limited support and is saturating (contrary to static_cast).

I'm not sure about other architectures, I believe arm does not have it. Still this could be a nice boost.