fix: avoid flushing a missing capture stream when tracking URLs

[cpruijsen]

Description

Stop the external-program tracking subprocess from calling flush() on a missing capture file at EOF. The iterator sentinel is now b"" (binary PIPE EOF) instead of "", and flush runs only if a capture file exists.

ExternalProgramTask tees stdout/stderr through a PIPE so it can scrape a tracking URL. Writes to the capture TemporaryFile were already skipped when capture_output=False. The empty-read branch still called file_to_write.flush(), which raises AttributeError: 'NoneType' object has no attribute 'flush' in the child process. Spark tasks always enable this tracker, so the traceback shows up on yarn/spark-submit runs even though it does not change the task result.

Fixes #3131

Decision: treat EOF as the end of the tracker and skip flush when there is no capture file.
Alternative: only guard the existing flush() and leave the "" sentinel / sleep loop.
Why: the unguarded flush is what the issue reports, and the else branch only ran because "" never matches b"". Can switch to the one-line guard if that is preferred.

Motivation and Context

#3131

PySparkTask on YARN logs file_to_write.flush() on None from _track_url_by_pattern. The reporter already noted it does not change the job result; the child exception is still noise on every tracked run without a capture file.

Have you tested this? If so, how?

I have included unit tests.

test_tracking_process_exits_cleanly_when_capture_output_disabled fails without the source change (AssertionError: 1 != 0, child traceback is the issue's AttributeError) and passes with it (exitcode == 0). Existing tracking tests in test/contrib/external_program_test.py and the Spark client/cluster tracking tests still pass.