Ci (compile and test in a GitHub action)

[JamesParrott]

This PR provides a basic working Continuous Integration framework, that can run the tests on every commit.

It builds on the compile (#24) and testing (#22) ones, but may require a minor merge conflict with either to be resolved, it having refactored the test running part into a reusable Action, so that the compiler workflow can use the same code.

However one of the tests currently fails unpredictably (when sDNA does not reproduce the exact same ERRORs as in correctout.txt). So until that issue (#21) is resolved (either by understanding the reasons behind the behaviour, or tweaking the tests to accomodate it), it would cause lots of frustration to rely on this workflow to
break commits, and stop releases, due to the high percentage of non-useful test failures.

Therefore, for now, the on push and on pull_request triggers been commented out:
.github\workflows\compile_and_test.yml

on:
  workflow_dispatch:
  # push:
  #   branches: [ "main" ]

  # pull_request:
  #   branches: [ "main" ]

The Pytest docs have great advice on dealing with flakey tests https://docs.pytest.org/en/8.0.x/explanation/flaky.html.

Unfortunately for sDNA, none of the Pytest plug-ins suggested by support Python 2 (flaky, pytest-flakefinder, pytest-rerunfailures, and pytest-replay:).

[JamesParrott]

In order to give confidence that the code in this PR is correct, in commit bc6def0 I have added a work around for issue #21 .

If the env variable ALLOW_NEGATIVE_FORMULA_ERROR_ON_ANY_LINK_PRESENT is Truthy (i.e. non-empty - don't use 0 as in Python bool("0") is True), extra code runs which allows lines in sDNA\sdna_vs2008\tests\correctout.txt to be skipped, only if:
a) both actual and expected lines start with: ERROR: Formula evaluation gave negative result for link n (in order to allow the expected n, and actual n to be diffferent). And:
b) in the actual output lines, the link numbers n in all error lines above, subequently occur in another line of the form: Polyline n id=i for some i (before the next negative formula error line, as above).

The tests pass locally in Python 3.12, and 2.7 using the command (in a venv with pytest and numpy installed):

set DONT_TEST_N_LINK_SUBSYSTEMS_ORDER=1 & set ALLOW_NEGATIVE_FORMULA_ERROR_ON_ANY_LINK_PRESENT=1 & pytest -rA

and reliably now in CI on Python 2.7 and on Pythons 3.5, 3.6, ..., 3.12 https://github.com/JamesParrott/sdna_plus/actions/runs/8215755710/job/22469718739

Filtering out this and any other strange behaviour from a testing strategy is sub-optimal. But they are only errors if the original test data (e.g. correctout.txt) is assumed to be correct or desired behaviour. The differences occur due to subtle reasons and are a lot of work to even understand, let alone fix.

Apart from these issues, 10s of thousands of other output lines are reproduced exactly, without being skipped. Therefore in the mean time, the benefits of automating the tests, and having a working (otherwise correct) non-flakey CI pipeline at all, are far more beneficial, than having no CI pipeline (or than having a CI pipeline with false negatives, if it is required to reproduce exactly the same errors, and the same output lines, but also in exactly the same order, as some previous build, in some other test environment did).