Builders: Implement non-host endianess writes using a single write.

[AndreasPK]

We do this by swapping the byte order in memory before we write.

This speeds up int64BE by about a factor of three on my machine (skylake).

int16BE hardly changes, i guess the write buffer is not a bottle neck for 16bit words. But beyond that it makes a large difference.

Before

    int16BE (10000):                                          OK (0.54s)
      7.68 μs ± 215 ns
    int32BE (10000):                                          OK (0.43s)
      11.6 μs ± 342 ns
    int64BE (10000):                                          OK (0.42s)
      22.7 μs ± 669 ns
    word16BE (10000):                                         OK (0.55s)
      7.68 μs ± 218 ns
    word32BE (10000):                                         OK (0.42s)
      11.6 μs ± 355 ns
    word64BE (10000):                                         OK (0.42s)
      22.7 μs ± 728 ns
    floatBE (10000):                                          OK (0.52s)
      14.6 μs ± 392 ns
    doubleBE (10000):                                         OK (0.47s)
      25.7 μs ± 689 ns

After

    int16BE (10000):                                          OK (0.53s)
      7.69 μs ± 223 ns
    int32BE (10000):                                          OK (0.30s)
      7.84 μs ± 397 ns
    int64BE (10000):                                          OK (0.29s)
      7.20 μs ± 590 ns
    word16BE (10000):                                         OK (0.55s)
      7.67 μs ± 217 ns
    word32BE (10000):                                         OK (0.30s)
      7.83 μs ± 393 ns
    word64BE (10000):                                         OK (0.29s)
      7.21 μs ± 595 ns
    floatBE (10000):                                          OK (0.34s)
      8.98 μs ± 328 ns
    doubleBE (10000):                                         OK (0.23s)
      10.8 μs ± 664 ns

[sjakobi]

Cheers! :)