Keep communication pipes between lsp-test and ghcide alive for the entire test

[crtschin]

Closes #4884.

Because the pipe is split up into separate handles and sent off to the separate component calls to lsp-test (async through ghcide in in-thread to lsp-test directly to communicate the test messages), a GC when either is in the process of being shutdown, will close one end of the pipe. This forms a race condition where the other end of the pipe hasn't been terminated and now unexpectedly reads a EOF.

For a more comprehensive explanation see my previous attempt at fixing this via lsp-test by being lenient on shutdown. I think this PR is nicer compared to that one, as it tackles the core issue, the handles being GC'd prematurely.

[crtschin]

So on the plus side, I don't see errors related to a closed handle anymore (no Language server unexpectedly terminated or failed hPutBuf: resource vanished). But the tests are still flaky with lsp message timeouts and errors related to progress reporting. Those are documented here though. Calling this a win, and leaving the remaining flaky failures for a follow up.

[soulomoon]

@crtschin Thanks for trying to improve the flaky test situation.
Regarding to progress report problem, here is some context might be helpful #4714

[fendor]

LGTM!
In the comment, could you explain where the read is happening?
It doesn't look trivial.