Add recipe for tuning distributed calculations

[RizzoHou]

The parallel/ chapter currently covers how many processors to use, but the
"Distributed memory parallelism" part of benchmarking.rst is a placeholder,
and introduction.rst notes that thread affinity settings are "beyond the
scope of this tutorial". This adds a page covering that ground: the settings
that decide how well a given number of processors is actually used.

New page parallel/tuning.rst, covering

• reading the internal timers first, so that effort goes where the time is,
• solver choice, including why a linear scaling solver is not automatically
  faster below its crossover size, and why ELPA's autotuner can cost more
  than it saves for a short SCC cycle,
• splitting node cores between MPI ranks and OpenMP threads, and the
  UseOmpThreads safety net,
• process and thread pinning, with two silent traps: setting
  OMP_NUM_THREADS inline rather than exporting it before the launcher, so
  that ranks get pinned to a single core while the code still believes it
  has several threads; and asynchronous progress threads being spawned per
  rank rather than per node, which deadlocks any layout that already fills
  every core,
• Parallel { Blacs { BlockSize } }, which is an input-file-only experiment,
• ELPA's own kernel selection environment variables,
• forcing MPI collective algorithms, and checking the fabric selection,
• a scanning method, including measuring the run-to-run scatter first and
  checking that the SCC iteration count and total energy are unchanged before
  believing a timing difference.

The numbers quoted are measurements on a 16,000 atom SiC supercell over four
nodes of 2 x 28-core Xeon 8280 with Intel MPI and MKL, made while preparing an
ISC student cluster competition entry. They are presented as indications of the
size of each effect rather than as recommended values, since every one of them
depends on machine, libraries and problem.

Two small edits to existing files: a label on benchmarking.rst so the new
page can link to it, and a pointer to the new page from the placeholder
section. The topic list in that placeholder is left alone, since the remaining
items (scaling examples, Groups, latency) are not covered here.

Built locally with Sphinx, no new warnings.