fix: create consistent output across architectures

[nlf]

thanks to architecture specific optimizations in the default deflate compression strategies, we end up with inconsistent output between arm64 and x86_64.

this change forces the compression strategy to one that appears to give more consistent results. they are larger, but since they're just temp files that feels like an acceptable trade off in the name of consistency.

for posterity's sake i'll mention here that Z_HUFFMAN_ONLY also provides a consistent end result, though i found better performance and compression with Z_RLE in some brief testing of a handful of npm related repositories.

note: i updated the CI matrix to match that of https://github.com/npm/cli

yes, the windows tests fail. however, i felt it was important to test the updated compression strategy for consistency in windows as well as linux and osx. i can fix the rest of the tests for windows before this lands, or we can land it as-is with a follow up PR to fix the tests for windows

edit: the node 10.1 test failures were a surprise, though appear to be unrelated to these changes

References

Closes npm/cli#2846

[isaacs]

Hmm... reproducible builds are good, but I feel like sending considerably more data over the wire is not great.

[isaacs]

Is it any less bad if you add

{
  windowBits: 15,
  memLevel: 9,
  level: 9,
  strategy: 3,
}

[nlf]

I'm confused, is the dir fetcher here in pacote also what libnpmpack uses to create the tarball for upload?

[nlf]

those settings make no difference at all.

also, confirmed that yes this is indeed the module that packs a tarball for publishing. so, yes, you're right, this is not what we want to do by default. i can tweak this so that this setting will only take place as part of fetching and packing a remote git repository.

[nlf]

closing to rework this and split it into two distinct pull requests