Add Copilot instructions for PR review

[kandrosov]

Why

This repository had no copilot-instructions.md, so Copilot reviewed its PRs with no idea what
this code is for or what goes wrong in it.

The file added here is written for PR review, not onboarding — the five existing files in the
ecosystem were generated by Copilot's repository-onboarding feature in November 2025, and are
being rewritten in the same way (cms-flaf/FLAF#310 and its companions): what a useful comment
looks like here, the invariants worth checking with the failure signature each produces, the rule
that documentation ships in the same PR, and — deliberately — what not to comment on, since
noise is an automated reviewer's main cost.

Framework-wide rules are not duplicated; the file points at FLAF/.github/copilot-instructions.md
for them.

Testing

Documentation only — no code change. Every path, filename and directory referenced was verified to
exist in the current tree, and the facts table is dated so the next reader knows how much to trust
it.

What it says

PlotKit is the one repository in the ecosystem that is an ordinary installable Python package with
unit tests that run anywhere — no CVMFS, no grid proxy, no CMSSW. So "we cannot test this"
does not apply, and asking for a test alongside a change to spec.py, config.py,
histogram.py, rootcompat.py or a plotter is a legitimate review comment. That is stated
explicitly.

The rest is what a screenshot would not reveal:

• Backend parity — a feature added to the matplotlib/mplhep path but not the ROOT/cmsstyle one
  produces plots that differ by backend, silently, because both succeed.
• Specs are a contract — PlotKit is consumed by FLAF transparently, so renaming a key or
  changing a default breaks three analyses; silent default changes surface months later in a plot
  that looks subtly different.
• ROOT is optional — rootcompat.py exists so the package works without it; a module-scope
  ROOT import breaks the pure-matplotlib install.
• Correctness a plot hides: dropped overflow/underflow, bin edges assumed uniform, ratio error
  propagation, normalisation applied twice, a stack ordered by dictionary iteration.