(WIP) sytrd use gemv instead of symv

[EdDAzevedo]

Code modifications to sytrd() and latrd() to use gemv() (general matrix vector multiply) instead of symv() (symmetric matrix vector multiply).

In some implementations and problem sizes, gemv() may give higher performance compared to symv(), even though symv() should perform only half the work and touch about half the data. The implementation of symv() might also be using atomic update operations.

The changes include:

1. allocating more work storage in sytrd() to store the conceptually untouched strictly upper triangular or strictly lower triangular part.
2. invoke kernels in sytrd() to save (on entry) and restore (on exit) the triangular parts.
3. invoke kernels in latrd() (on entry) to copy the strictly lower triangular part or strictly upper triangular part of matrix to enforce symmetry. This is to allow gemv() to replace calls to symv().
4. modified xxTRD_BLOCKSIZE from 32 to 64 to reduce the cost for matrix copy for enforcing matrix symmetry in latrd()

On gfx1030, using rocsolver-bench -f sytrd --precision s --iters 5

n	using gemv (us)	using symv (us)
1024	62,201	73,580
2048	137,507	190,302
4096	335,996	522,818
8192	2,161,237	1,925,926

On MI300 (splinter-126-wr-d3, gfx942)

n	using gemv (us)	using symv (us)
1024	40,310	53,154
2048	94,027	157,111
4096	237,926	484,525
8192	689,551	1,683,223