arm64: fix the SME2 sbgemm kernel's ZA handoff and BFMOPA predicates

Two fixes from Michael Tesch's M4 Pro testing:

- s2b_blk had its own __arm_new("za"), so the call from the equally-
  annotated s2b_drive set up a lazy ZA save that __arm_tpidr2_save
  (OpenBLAS's aborting weak stub) committed. It now shares ZA from the
  drive function through S2_S S2_ZA like the fp32/fp64 block functions.
- BFMOPA governs per 16-bit element, but the predicates were
  svptrue_b32(), which enables only the even bf16 lanes and drops the
  k+1 half of every pair (every k >= 2 wrong, k = 1 right). Now
  svptrue_b16().

Also apply the suggested thread cap to SBGEMM: sme_level3_threads()
capped only real precision (!BFLOAT16), so default threads contended
for the SME units - 839 GFLOPS at n = 4096 against about 1190 on one
thread and 2347 with one thread per unit on the M4 Pro. The condition
is now (!BFLOAT16 || !BGEMM): bgemm keeps the NEON kernels and its
threading.

With the fixes his run reports the sbgemm/BLAS/CBLAS/utest suites
passing and 3000 randomized trials clean; 1193 GFLOPS at 2048^2 single
thread against 0.48 on develop.

Author: 0tomhoyt

common_thread.h

@@ -157,7 +157,7 @@ extern int blas_omp_threads_local;
 extern int blas_is_num_threads_set_explicitly;
 
 #if defined(ARCH_ARM64) && defined(OS_DARWIN) && defined(__clang__) && (defined(VORTEXM4) || defined(DYNAMIC_ARCH)) && \
-    !defined(COMPLEX) && !defined(BFLOAT16) && !defined(HFLOAT16)
+    !defined(COMPLEX) && (!defined(BFLOAT16) || !defined(BGEMM)) && !defined(HFLOAT16)
 #include <sys/sysctl.h>
 #ifdef DYNAMIC_ARCH
 extern char *gotoblas_corename(void);

kernel/arm64/sbgemm_kernel_sme2.c

@@ -92,7 +92,7 @@ S2_INL void s2b_op(const int SUB, int t, svbool_t p, svbfloat16_t b, svbfloat16_
 /* ZA += or -= the A panel (wm rows) times the B panel (wn columns) over k steps of the packed layout. */
 S2_NOINL void s2b_mac(const int SUB, BLASLONG k, const IFLOAT *a, BLASLONG wm, const IFLOAT *b, BLASLONG wn) S2_S
     S2_ZA {
-  const svbool_t pt = svptrue_b32();
+  const svbool_t pt = svptrue_b16(); /* BFMOPA governs per 16-bit element */
   const svbool_t pm0 = svwhilelt_b16_s32(0, (int32_t)wm);
   const svbool_t pm1 = svwhilelt_b16_s32(S2B_UN * 2, (int32_t)wm);
   const svbool_t pn = svwhilelt_b16_s32(0, (int32_t)wn);
@@ -116,9 +116,8 @@ S2_NOINL void s2b_mac(const int SUB, BLASLONG k, const IFLOAT *a, BLASLONG wm, c
 
 /* Block of wm rows and wn columns of C; sgn 1 for C += alpha AB with alpha = +-1 (the AB part
    added or subtracted in ZA), sgn 0 for a general alpha scaled on the way out of ZA. */
-__arm_new("za") __arm_locally_streaming static void s2b_blk(BLASLONG k, const IFLOAT *a, BLASLONG wm, const IFLOAT *b,
-                                                            BLASLONG wn, FLOAT *C, BLASLONG ldc, FLOAT alpha,
-                                                            int sgn) {
+S2_NOINL static void s2b_blk(BLASLONG k, const IFLOAT *a, BLASLONG wm, const IFLOAT *b, BLASLONG wn, FLOAT *C,
+                             BLASLONG ldc, FLOAT alpha, int sgn) S2_S S2_ZA {
   if (sgn) {
     s2_c_load(1, S2_NT, C, ldc, (int)wn, (int)wm, S2_LOAD, (FLOAT)1);
     s2b_mac(alpha == (FLOAT)-1, k, a, wm, b, wn);