Note the mechanism used needs expanding to all other record types.

Functions passed to the multiprocess module must be picklable, which
means they must be module-level functions.
Also adding the "coding" comment to let me use unicode characters

This reverts commit da1dca9.

Now that Python2 support is dropped, Python3 strings are unicode anyway

Question is asked in the PR instead.

None (and ctypes) will now raise exceptions if used as a value.
Added tests for rejecting None at all stages of processing.

"None" is considered a valid value for String records - it'll result in
the literal string "None" appearing in the output record, which is
acceptable behaviour.

About to introduce a whole new swathe of tests using caget/caput, and so
extra organisation was necessary

Also tweak test names for to make it easier to distinguish them

This includes many fixes for handling different types through CAGet and
CAPut, as values passed or received need a fair amount of massaging to
align with existing test framework

cothread isn't on Windows, but we still want to  run some tests there

MacOS CI test runners seem very slow and have intermittent
timeout-related failures

This should have been done as part of the removal of Python2 support.

This required a large amount of checks in many places to cover edge
cases. I think this test code is probably getting too complex and
needs simplification.

It appears newer cibuildwheel versions try to build for more platforms.
epicscorelibs doesn't support musllinux, so skip building there.

This new model will have the parent process wait until IocInit() has
successfully completed before continuing with the test

The child process had need to do a queue.put() followed immediately by a
queue.get(). On most platforms the main process would interuupt by
pulling from the queue, but this was not guaranteed. To solve this using
queues would have required a second queue for signalling.

We work around this by using Pipes instead. These provide the
read-write semantics we want - the ability to wait for an explicit message
from the parent process - without needing a side-channel for more
signalling.

Hopefully this will fix the MacOS issues where it seemed caget was
returning values from a previous iteration of the tests, possibly
because the test process was running much faster than the child process
could keep up with.

Need to investigate issues with Windows no longer running tests properly
on cibuildwheel 2+ - the tests fail to find the softioc/iocStatsDb dir
as it is no longer packaged on Windows

This reverts commit db85ca3.

This reverts commit bf47f63.

This is an attempt to get Windows softlinks built into the wheel

Should not have removed it in the previous commit!

The fix is to pin cibuildwheel at 2.2.2, but I'm doing this for the
failing build log so I can raise an issue against cibuildwheel

This appears to resolve issue regarding Windows symlinks not copying.
See pypa/cibuildwheel#962

Previously it would run twice, for no good reason.
This also adds running a weekly build. This may or may not be necessary,
but it's part of the new template module so we'll add it for consistency

Pytest will now output JUnit-format XML results, which the action should
read and display on the PR in a nice format.

cibuildwheel 2.3 upgraded the version of pip it uses, but that version
altered how symlinks were handled on Windows. Specifying this
environment variable should restore the behaviour for now. It'll be
deprecated eventually though so we will need to do something else
long term.

The new model of tests create a lot of records in each subprocess,
rather than one subprocess per record. This hugely decreases runtime -
down to 0:30 from 4:30 runtime on my machine.

Unfortunately this mechanism does reduce resolution of results, as
they'll fail on the first issue rather than finding all issues.

We don't run this formatter anywhere automatically, this was done
manually.

Specifying it in the setup.cfg file didn't seem to work, so try in line

Suggested workaround of activating out-of-tree-build seems to not work
on python 3.8+ and so it isn't suitable

This approach specifies the environment variable inside the
CIBW_ENVIRONMENT variable, which will mean it won't interfere with
cibuildwheel setting up its own environment

The goal here is to handle the conversion between Python and EPICS value
representation in the most consistent way possible.  Each record type has
an underlying representation corresponding to what is written to EPICS, but
the user interface continues to work with standard the Python
representation.

Also destroy readable layout to avoid flake8 test failure.

It seems numpy has stopped publishing manylinux1 builds, in favour of
manylinux2014. So we must update to follow (as it's a right pain to build
numpy ourselves - the CI runners can't do it at the moment)

Handle null termination properly and ensure we have a proper copy of the
buffer.

Also a bit of tidying of device.py

…tests

These were developed on the "sort_settings_values" branch but were
never committed there. Still valid test cases, although they require a
little refactoring to fit with the changes made on this branch.

Also neaten names for Validate tests, and tweak None tests to add
extra exception

The tests for LongStringOut records are broken.
Validate is never called - caput doesn't even work.

Discussed with Michael.
Probably will be more tests to delete, pending further discussion.

Changes made after discussion with Michael regarding what is and is not
a valid test scenario, with Michael's latest changes to record value
handling.
Note there are pending changes to the record code itself. This is just
test code changes.

Ensure that the waveform type and length are specified during record
creation to avoid problems with .set() later on.