Bump Win32

Bump filepath upper bound to allow 1.5

Remove tests/ subdirectory

The tests in the `tests` subdirectory are only run by the GHC testsuite.
Following ghc/ghc#22622 the tests
will be moved from the git submodule to the `testsuite/tests` directory
in GHC itself.

Add support for the GHC JavaScript backend (node.js)

Prevent a segfault in createProcess on Mac

Closes #295

Bump base bound to <4.20

For GHC 9.8.

This adds support for the GHC JavaScript backend. The resulting code can
be run on node.js.

Update tests for the new test way `ghci-opt`

Tracking ticket: #23059

testsuite: Remove calls to extra_clean

It has been a no-op since GHC commit 1a9ae4b39c057d2,
which is present in GHC 8.2.

Revert "Use POSIX_SPAWN_CLOEXEC_DEFAULT in posix_spawn.c when available"

This reverts commit f47c8699.

testsuite: mark tests as req_process

cbits/fork_exec: Eliminate potential leakage of comm. pipes

Previously we used `fork` followed by `fcntl` to create a pipe with
`O_CLOEXEC`. However, this is potentially racy as another thread may
`fork` in between in these two calls, resulting in the pipe of the
former thread being available to the child of the latter.

Fix this by using `O_CLOEXEC` when creating the communications pipe.

Closes #287.

Use POSIX_SPAWN_CLOEXEC_DEFAULT in posix_spawn.c when available

Address #284 and #283

Closes #284.

Previoussly we would call stopTimer sooner than necessary, then fail to
restart the timer if find_exectuable failed. Avoid this by deferring the
stopTimer call until we are just about to fork.

Closes #283.

cbits: Ensure that fork pipe doesn't shadow std fds 

Closing an already closed FD is OK and should not cause failures.

Previously we would assume that the pipe used to communicate
errors from the child back to the parent did not shadow the standard
descriptors (stdin, stdout, and stderr). Somewhat surprisingly, this
appears to hold most platforms. However, OpenBSD appears to be a notable
exception. This lead to the failure of the `processT251` test.

Avoid relying on this assumption by `dup`ing the pipe fds until they end
up out of the standard fd range.

Closes #266.