dstedc, sstedc: remove dead ICOMPZ.EQ.0 branch

[jschueller]

The IF( ICOMPZ.EQ.0 ) block calling DLASRT/SLASRT quick sort after the endwhile loop is unreachable -- ICOMPZ.EQ.0 already exits via CALL DSTERF/CALL SSTERF + GO TO 50 earlier. Remove the dead branch and its external declaration, keeping only the selection sort path.

Closes #1143

[langou]

Argh. I do not like it.

Yes, this fixes a dead branch for the current code.

But the current code reads

lapack/SRC/dstedc.f

Lines 297 to 312 in 693f114

	*
	* If the following conditional clause is removed, then the routine
	* will use the Divide and Conquer routine to compute only the
	* eigenvalues, which requires (3N + 3N**2) real workspace and
	* (2 + 5N + 2N lg(N)) integer workspace.
	* Since on many architectures DSTERF is much faster than any other
	* algorithm for finding eigenvalues only, it is used here
	* as the default. If the conditional clause is removed, then
	* information on the size of workspace needs to be changed.
	*
	* If COMPZ = 'N', use DSTERF to compute the eigenvalues.
	*
	IF( ICOMPZ.EQ.0 ) THEN
	CALL DSTERF( N, D, E, INFO )
	GO TO 50
	END IF

And so the idea is that someone could comment out

      IF( ICOMPZ.EQ.0 ) THEN
         CALL DSTERF( N, D, E, INFO )
         GO TO 50
      END IF

so as to use the divide and conquer symmetric tridiagonal eigensolver to compute the eigenvalue only of the matrix instead of QR/QL symmetric tridiagonal eigensolver (STERF).

Why would someone want to do this? (1) maybe D&C is faster than STERF for eigenvalue only, (2) maybe the user wants to get the bit-by-bit same eigenvalue when calling STEDC with COMPZ = 'N' or COMPZ = 'I' or COMPZ = 'V'.

Related, note the recent work of Ruiyi Zhan and Shaoshuai Zhan on accelerating D&C when only eigenvalues are desired: https://arxiv.org/abs/2605.26599

I'd almost advocate to comment out

      IF( ICOMPZ.EQ.0 ) THEN
         CALL DSTERF( N, D, E, INFO )
         GO TO 50
      END IF

and have D&C stay true to D&C. Even if it is slower for eigenvalue computation, that's what it is. I want to say to STEDC "Be yourself! Be proud of you are! Do not try to be a QR algorithm".

Note: It'd be interesting to see if this passes the test suite since D&C for eigenvalue only is not really tested currently.

All in all, I do not support this PR.