Quick return with identity Houseolder transformation if there are no off-band elements to annihilate

[RMeli]

If the input matrix is banded and has a band size smaller than the target band size for the reduction, nans are produced due to a division by zero when defining tau:

DLA-Future/include/dlaf/eigensolver/reduction_to_band/impl.h

Lines 102 to 105 in 1fd315b

	const T norm = std::sqrt(x0_and_squares[1]);
	const T x0 = x0_and_squares[0];
	const T y = std::signbit(std::real(x0_and_squares[0])) ? norm : -norm;
	const T tau = (y - x0) / y;

This PR introduces an early termination when x0_and_squares[1] == 0 by returning tau = 0, thus circumventing the problem. Fix #974. Thanks @albestro for the help in identifying the issue.

---

With this PR, the situation in CP2K is the following (using the DLA-Future eigensolver for every matrix of size 2x2 or higher):

------------------------------- Summary --------------------------------
Number of FAILED  tests 2
Number of WRONG   tests 0
Number of CORRECT tests 2935
Total number of   tests 2937

Summary: correct: 2935 / 2937; failed: 2; 28min
Status: FAILED

*************************** Testing ended ******************************

All regression tests still returning nans (see #974) pass. Of the two remaining tests one has been fixed in CP2K (missing DLAF/pika initialization when using CP2K as a library), while the other one also fails with ScaLAPACK.

[RMeli]

cscs-ci run

[RMeli]

cscs-ci run

[RMeli]

cscs-ci run

[RMeli]

cscs-ci run

[RMeli]

Current CI failures should be fixed by #983.

[RMeli]

cscs-ci run

[RMeli]

cscs-ci run

[RMeli]

cscs-ci run

[RMeli]

cscs-ci run

[RMeli]

Issue reproduced in CI.

[RMeli]

cscs-ci run

[RMeli]

cscs-ci run