Add a coverage test job with a compiled Cython package to compare plain Python and compiled Cython both for completeness and speed.

Exclude end-to-end tests from test runs with '--cython-only' since they execute arbitrary test code.

This commit fixes a crash of Cython when generator expressions are used in cdef functions in pure python mode.
Closes https://github.com/cython/cython/issues/3477

The implementation of __Pyx__PyNumber_Float() wasn't checking whether
errors occurred on the calls to __Pyx_PyUnicode_AsDouble(),
__Pyx_PyBites_AsDouble(), and __Pyx_PyByteArray_AsDouble(). Errors from
those functions were now properly propagated to the caller.

Closes https://github.com/cython/cython/issues/4143

Closes https://github.com/cython/cython/issues/1851

Disable some unicode fstring tests for the debugging (f"{...=}") feature since unicode strings come out as "u'...'" in Py2.

Closes https://github.com/cython/cython/issues/3796

Fixes https://github.com/cython/cython/issues/3954

The problem seems to be that __Pyx_CallUnboundCMethod0 optimizes the call incorrectly. This patch avoids it getting sent to that mechanism.

Don't access directly PyObject.ob_type and PyObject.ob_refcnt members.

Correct the return type of the ..._GET_TZINFO() macros (which return a borrowed reference) and add corresponding properties so people don't have to use them.

https://foss.heptapod.net/pypy/pypy/-/issues/3404

Let a tracing test run slower to prevent div-by-zero crashes in line_profiler due to measured zero runtime.

Reconfigure ccache and add the backend (C/C++) to the cache key to distinguish between the compiler jobs.

Make sure that test failures are easy to find in the logs by putting all information into the assertion.

Difficult to say which is best, but something around 7-9 apparently.
Accumulated C/C++ compile times only ever go up with increasing shard count (tried 5, 7, 8, 9, 12) but the total test time is lowest at 7-9, depending on the specific run. So, let's waste less cycles overall.