ProcessedContainerSettings.from_container_settings() rebuilds the processed
container settings from the raw ContainerSettings, but it never copied the
enable_profiling field, so the processed settings always used the dataclass
default of False. run_container_singularity() reads the flag off the processed
settings (config.enable_profiling), so the profiling branch never ran: no peer
cgroup was created, cgroup_wrapper.sh never executed, and no usage-profile.tsv
was written, regardless of the user's containers.enable_profiling config value.

This regressed in Reed-CompBio#390, which moved enable_profiling from a top-level config
field to the nested container settings and added the field to both
ContainerSettings and ProcessedContainerSettings, but missed wiring it through
the conversion in from_container_settings().

Add the missing assignment, plus a regression test in test_config.py asserting
that enable_profiling survives the full config-parsing path.

subprocess.run rejects capture_output=True together with an explicit
stderr argument (capture_output is shorthand for stdout=PIPE, stderr=PIPE).
In the profiling/cgroup branch of run_container_singularity this raised
"ValueError: stdout and stderr arguments may not be used with
capture_output." on every Apptainer run with enable_profiling=true.

Set stdout=PIPE and stderr=STDOUT explicitly so the container's stderr is
merged into the captured stdout, preserving the original intent of
capturing combined output via proc.stdout.