This adds build that runs expensive tests, including Vc on with at least AVX.
The build is allowed to fail to compensate for the possibility of no runner
beeing available.

Feature/float128

See merge request !518

[bugfix] add missing include for `std::integer_sequence`

See merge request !521

[bugfix] VariableSizeCommunicator fixes

See merge request !520

Otherwise we are trying to receive more requests than are there and get
dubious exception because of resizing vectors with MPI_UNDEFINED.

MPI_Wait* and MPI_Test* will ignore these.

Indeed we cannot assume that we always have to send data along one.

[SIMD/test] Do not access indeterminate simd vectors with lane().

See merge request !519

Depending on implementation, this may require bitmask operations on the
indeterminate value, which requires an lvalue-to-rvalue conversion, which
results in undefined behaviour except in some very special cases.

Add helper function unpackIntegerSequence

See merge request !495

[SIMD] Test the index vectors the same way we test other vectors.

Closes #125

See merge request !516

In the tests for operations we check whether an operation is supported on
scalars, and if so, we also test is in simd types.  This suppresses tests for
e.g. `%` on simd types made up of floating point types.  Before this patch we
sometimes did the check on proxies instead of scalars.  This could lead to
tests be incorrectly not suppressed.

This had been watered down because vectorclass does not implement some
functionality for integer vector types (like certain forms of division or
bitshift).

Instead of watering down the test, we're now going to provide that
functionality in a poor efficiency version.

[SIMD] Get rid of IsStandard and all specializations.

Closes #122

See merge request !514

Fixes: #122.

A non-const class rvalue is, in fact, assignable, unless you take special
care.  In the tests we try to apply each operator to `Scalar<V>` in an
unevaluated context, and if that succeeds we require that it works for `V`
analogously to how it does for `Scalar<V>`.

That heuristic fails if `Scalar<V>` is a class (such as `std::complex`),
because assignment works even to rvalues (usually), but trying to apply the
assignment to a `V` rvalue and then capturing and examining the result will
lead to undefined behaviour.

[AlignedNumber] Run the SIMD testsuite.

Closes #123

See merge request !511

[ArithmeticTestSuite] Inherit constructors from TestSuite.

See merge request !512

This makes it possible to e.g. attach a name to the testsuite object.

Also run the simd testsuite on AlignedNumber, sind it does provide a SIMD
abstraction.

Also fixes one issue with that abstraction (though I do not recall why exactly
I did not allow `const Scalar<T> &lane(T &&)`, but required the return type to
be one of `Scalar<T>` or `Scalar<T>&&`).

[Simd::IsStandard] Disable for AlignedNumber too.

See merge request !510

This fixes the tests for istl again.

See !509, #122.

[Simd::IsStandard] Disarm.

Closes #121

See merge request !509

Fixes: #121.

Stop-gap measure to to accept any type in the adaption for standard types.

Minor moderization of pool allocator

See merge request !507

Also use a `std::vector` instead of plain `new[]` and `delete[]`.