Fix intermittent failure of test/leaky-handles.py

[bdbaddog]

The test counted the entries in /proc//fd using a shell pipeline. $$ is the shell, not ls, and for a pipeline the shell creates a pipe, forks both commands, and only then closes its own copies of the pipe ends. When ls reached getdents() first it saw those too and printed 4 or 5 instead of 3, failing the exact stdout match. Measured at roughly 0.5% when the shell and ls contend for a cpu (2 failures in 400 runs of the test pinned to one cpu), which is why it showed up under parallel and loaded test runs but never in isolation.

The forked command now reports the inherited handles itself, so no shell pipe is involved: 0 failures in 400 runs under the same conditions.

This also restores what the test is meant to check. The handle it leaked could not reach the child for two unrelated reasons: the 'with' block closed it when the SConstruct finished being read, well before the build phase forked the command, and handles are close-on-exec by default since Python 3.4 (PEP 446) regardless of what SCons does. Binding the file object to an SConscript global is not enough either, as it is garbage collected before the build phase; the test now keeps an os-level handle open and marks it inheritable. Verified it fails as intended when close_fds is removed from SCons/Platform/posix.py.

Contributor Checklist:

• I have created a new test or updated the unit tests to cover the new/changed functionality.
• I have updated CHANGES.txt and RELEASE.txt (and read the README.rst).
• I have updated the appropriate documentation