[TheNetherWatcher]

Closes #464.

wr_dat_file packed formats 16, 24 and 32 by masking every sample into a separate int64 array for each byte, interleaving those arrays, and then casting to uint8. That uses about 4 bytes of memory for every byte written, plus an up-front copy of the whole signal.

As @bemoody suggested in #464, this now converts the signal to little-endian 16 or 32 bit integers and views the result as bytes (dropping the top byte of each sample for format 24), like format 80 already does. The explicit < keeps the output little-endian on any machine.

The up-front copy is gone too. It was only there because formats 80 and 212 changed the array in place, so those two now compute their offsets out of place instead. The caller's array is still never modified.

Peak memory and time of wr_dat_file for 4 hours × 8 channels at 500 Hz (460 MB of int64 samples), best of 3:

Format	Time before → after	Peak memory before → after
16	331 → 25 ms	1978 → 110 MiB
24	613 → 188 ms	3021 → 385 MiB
32	840 → 56 ms	3955 → 220 MiB
80	60 → 55 ms	494 → 494 MiB
212	620 → 251 ms	961 → 961 MiB

The written files are unchanged. The new TestWrDatFile tests compare the bytes for every writable format against a sample-by-sample reference (int.to_bytes and the format 212 bit layout) for odd and even lengths, the format's minimum and maximum values, Fortran-ordered or sliced input, and multi-frequency (expanded) signals. They also check that the input array isn't modified. They pass on main as well, and fail if the byte order or the dropped byte for format 24 is changed.

pytest -m "not network" passes (121 tests), and the docs build passes.

[tompollard]

Thanks, this looks great, and the tests against a sample-by-sample reference are a nice touch!

I checked it independently:

• The written bytes are identical to main for formats 80, 212, 16, 24, 32 and 516, with int16/int32/int64 input, C-ordered, Fortran-ordered and sliced arrays, odd and even lengths, and multi-frequency signals. The input array is never modified.
• Removing the up-front copy is safe: the FLAC branch only shifts arrays that astype has already copied.
• For a 1-hour, 8-channel signal at 500 Hz, peak memory went from 494 to 27 MiB for format 16, 755 to 96 MiB for format 24, and 989 to 55 MiB for format 32.