Measure benchmark memory usage natively instead of via /usr/bin/time

[guitargeek]

This supersedes #190 (itself a reincarnation of #95): same goal, but with a native implementation instead of shelling out to the time binary.

Motivation

The interpreter benchmarks obtained their RSS counter by prefixing the tutorial invocation with a hardcoded /usr/bin/time -v, parsing its output file with grep/awk, and passing the result through a tmp_mem_file in the working directory shared by all benchmarks. On systems without the time binary, the shell command failed before ever running the tutorial, so the benchmarks silently measured a failed shell invocation and reported RSS=0.

The PyROOT benchmarks were in even worse shape: no memory measurement at all (RSS was always 0), and the shell quoting of the TPython::Exec() call was broken, so the tutorials never actually ran — root only printed an interpreter parse error, which doesn't show up in its exit status either.

Changes

• Add RB::RunCommandMeasuringRss() to RBSupport: runs a command like std::system() and reports the maximum resident set size of the child process tree via wait4(). This is the same value that /usr/bin/time -v reports as "Maximum resident set size", but with no external binary, no temporary files and no output parsing.
• Use it in the interpreter and PyROOT benchmarks, and fail the benchmark with a clear error (SkipWithError) when a tutorial cannot be run, instead of silently reporting garbage timings.
• Fix the PyROOT TPython::Exec() quoting and propagate failures of the interpreted code into the exit status via gSystem->Exit().
• Fix the locations of the PyROOT tutorials after the tutorials reorganization in ROOT. Nine of them no longer exist and were dropped, as were geometry.py/na49view.py, which cannot run standalone (they need a geometry created by other macros — a failure the new error reporting surfaced).

Verification

Against a recent ROOT master build:

• Test_nopch_mlp_higgs went from a bogus 18 ms / RSS=0 (on a machine without /usr/bin/time) to a real 4.2 s / RSS≈262M.
• All six remaining PyROOT benchmarks now actually execute their tutorials (4.3–6.0 s each, RSS 444–490 MB), with failures reported as benchmark errors instead of fake numbers.

Note for the dashboards

Timings and RSS values for these benchmarks will jump discontinuously once this is merged: the old numbers were measuring a failing shell command, so the apparent "regression" on the graphs is the benchmarks starting to work.