Non temporal data transfers

[DiamonDinoia]

1. Adding stream API for non temporal data transfers
2. Adding xsimd::fence as a wrapper around std atomic for cache coherence
3. Adding tests

Draft because I need to double check the API levels ( i.e I am not using AVX2 functions in AVX and so on). I just wanted some feedback while I do the finishing touches.

[serge-sans-paille]

Some generic thoughts:

• I'm unsure the fence belongs to xsimd, but I like being proven wrong, maybe show us a code example that uses it?
• load_stream or stream_load or streaming_load?

On arm64, there's no support for non-temporal loads (https://developer.arm.com/documentation/100048/0100/level-1-memory-system/memory-prefetching/non-temporal-loads), the corresponding instruction do exist (LDNP/STNP) but I failed to find the related intrinsic.

There seems to be something equivalent in riscv (see riscv-non-isa/riscv-c-api-doc#47)

I couldn't find anything for webassembly nor Power. So that's quite a niche, but I'm fine with adding those though.

[DiamonDinoia]

1. I went for load_stream and store_stream so that it is consistent with [load|store]_[un]aligned... (Also load_non_temporal was too long and load_nta is not clear).
2. I added fence for convenience. I have no strong feelings on it. We can always think about adding it in the future. In the end on x86, I was recently made aware that it is not needed on a single core application. In parallel applications, atomic is likely to be imported anyway.
3. About ARM and RISK-V what about making our own intrinsics by wrapping the inline assembly? I sadly do not know about ARM all that much to be able to promise I will help

Cheers,
Marco

PS: sse2 adds APIS for non temporal stores of scalars of 32/64 bits. I am not sure the fit within xsimd though