Implement Eio.Path.native as a Pi backend and add Windows support

[talex5]

This is split off from @avsm's #908 and rebased on top of #913 and #915.

Incorporates the approach from #738.

I haven't reviewed it yet.

[avsm]

I think this is broken now on Windows due to the changes in step handling from #913; I'll fix it up first before you need to review.

[talex5]

I think this is broken now on Windows due to the changes in step handling from #913; I'll fix it up first before you need to review.

Should be fixed now (I spotted it too).

[avsm]

I've been reading up more on Windows paths and have a commit to slightly simplify the nt path parser, but also split out the tests. Sadly my Windows machine keeps rebooting for reasons unknown so I'll have to push it once I get back to it in person later today I hope :-)

[avsm]

With apologies for the code churn, I've moved to mdx for the windows tests in 1fae694 which is a bit clearer to read than the denser alcotest. I'm trying to get all the portable mdx working again on Windows, but this is just the fs bit. It also drops an alcotest dep on Eio as a result.

[avsm]

my CRLF is messed up, sorry! Can revert 1fae694 (but my Windows machine has just rebooted again!!)

[talex5]

No problem; I squashed them and converted the newlines in fs.md (I left the existing test files alone).

Hmm, that doesn't work (failed in CI). I've switched it back to CRLF.

[talex5]

I think we've hit realworldocaml/mdx#294 here.

Some git installations are configured to translate newlines, some do not. If they are translated, the LFs in this repository will be converted to CR LF and the tests will fail. Conversely this translation can also hide problems in the other direction by making CRs disappear.

I tried using .gitattributes to force it to lf, which didn't work. Then I forced it to crlf, that didn't work either.

[talex5]

Can revert 1fae694 (but my Windows machine has just rebooted again!!)

I partly reverted this, but moved the tests of Nt_path to the main tests/fs.md, next to the POSIX ones. That means they get tested on Linux and macOS at least. The remaining tests (that use Eio_windows.run) are still using Alcotest and only run on Windows.