* rows = 0 or cols = 0 is now a legal input and
  takes quick return path
* Follow BLAS/LAPACK convention that the leading
  dimensions must be at least 1.

* Clean up ZROTG

Fix division by zero in [z]rotg

* The check da != ZERO is no longer necessary since there
  is a special case ada == ZERO, where ada = |da|.
* Add the missing check c != ZERO before the division.

Note that with these two changes the long double code
follows the float/double version of the code.

The cases
[ c  s ] * [ 0      ] = [ |db_i| ]
[-s  c ]   [ i*db_i ]   [  0     ]
and
[ c  s ] * [ 0      ] = [ |db_r| ]
[-s  c ]   [ db_r   ]   [  0     ]
computed s incorrectly. To flip the entries of vector,
s should be conjg(db)/|db| and not conjg(db) / da,
where da == 0.0.

* Fix potential integer overflow

Sapphire Rapids sbgemm fix

* Fix return value for invalid info
* Add missing checks for ldA, ldB
* Use reference-LAPACK like checks (ie ld=0,nrows=0 is invalid)

Add a trivial GEMMT implementation based on a looped GEMV

Co-authored-by: Alexis <35051714+amontoison@users.noreply.github.com>

WIP casts and cleanups