Add support for Zstandard compression

[jcadduono]

While it might be very early in the stages of Zstandard's development, I'm stuck at work with nothing to do and wanted to learn C (everything is guesswork based on grep -rHni 'lz4' . and a few peeks at zstd libraries). :)
Perhaps merging could wait a while due to the early stage development of Zstandard, but I do think at this point the Zstd library functions used are unlikely to be changed. I did some testing on my uncompressed initrd (~10 MB) and some small files (~1300 b) and figured the results from the current revision of Zstd were very impressive for realtime packet compression. I might build a bigger testing environment for individual packet compression later on to better confirm my results, but for now, I give you this:

initrd:

unc:    : -         @ 10395136 b
lz4-1   : 0.026s    @  6299540 b
snappy  : 0.031s    @  6287148 b
lz4-9   : 0.270s    @  5261604 b
gz1     : 0.156s    @  4978336 b
zstd    : 0.037s    @  4969289 b
gz9     : 1.600s    @  4590729 b
xz1     : 0.706s    @  3693364 b
xz9e    : 3.851s    @  3248256 b

packet:

unc     : -         @ 1032 b
lz4-1   : 0.001s    @  609 b
lz4-9   : 0.001s    @  586 b
snappy  : 0.000s    @  579 b
xz1     : 0.001s    @  520 b
xz9e    : 0.014s    @  512 b
zstd    : 0.001s    @  506 b
gz1     : 0.001s    @  485 b
gz9     : 0.001s    @  462 b

As for the negative impact of this patch, the FSE library throws 384 extra bytes into the buffer space or so I was able to interpret from fse.c, then Zstd throws another 9 extra bytes on top of that. Yikes!
Perhaps it would be better to more dynamically get COMP_EXTRA_BUFFER(len) based on current compression method. :/ In any case, we're talking bytes so it's probably not too bad.
ex, in comp.h:

 * Zstandard: len + len/128 + 512 + 12  (ZSTD_compressBound(len))
 */
#define COMP_EXTRA_BUFFER(len) ((len)/8 + 512 + 12 + COMP_PREFIX_LEN)

[cron2]

Hi,

after a bit of discussion in the OpenVPN team, we have decided that we do not want to support a larger number of compression algorithms, and have removed snappy as well - so, for the time being, we only support LZO (for compatibility) and LZ4 (for modern deployments).

gert