Benchmark changes the current working directory if self.run_dir is
set befor each benchmark run.
lld does not work with print_status_on_exit.so so reading and removing
status is no only done if a status file is available.

If the loader calls build/exec the execvp from build/exec is handled
by the system loader.

use the included loader of the build glibc instead of LD_LIBRARY_PATH.
Using LD_LIBRARY_PATH probably fails because of incompatible system
loaders and newer glibc versions.
Then we have to reverse the positions of {measure cmd} and {cmd prefix}
in the exec chain because ld loads only ELF binaries and {measure cmd}
could be a script.

Calling the loader fails on non ELF executables. Exec, formerly run_cmd,
now takes two options: -l LD_LIBRARY_PATH, -p LD_PRELOAD.
-p sets LD_PRELOAD and -l LD_LIBRARY_PATH before executing the rest of argv.

glibc no longer uses cmd_prefix in favor of LD_LIBRARY_PATH.

Originally the structure of the executed cmd was
{measure cmd} {allocator cmd prefix} {cmd} with the parent environment
except LD_PRELOAD was modified for the whole command chain.
Unfortunatly perf causes segfaults with some allocators and measuring
allocators cmd prefixes doesnt seem fair. So the new cmd chain looks like:
{allocator cmd prefix} {measure cmd} run_cmd <LD_PRELOAD> {cmd} without
touching the environment in python.

run_cmd sets LD_PRELOAD to the value it received in argv[1] and executes
argv[2] with the rest of argv. This does also measure code not part of
the actual benchmark but in a equal manner and not only for some allocators.