Remove flush to zero from bf16

After closely analyzing Google Brain codebases, we decided that flushing
to zero was the wrong thing to do. Intel and AMD probably designed their
microprocessors to always flush to zero for the wrong reasons. It should
have been made conditional on FTZ being set in MXCSR like other opcodes.

See ggml-org/llama.cpp#7843

Author: jart

bf16.h

@@ -52,10 +52,6 @@ static inline float from_brain(uint16_t h) {
 
 /**
  * Converts float32 to brain16.
- *
- * This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
- * Subnormals shall be flushed to zero, and NANs will be quiet.
- * This code should vectorize nicely if using modern compilers.
  */
 static inline uint16_t to_brain(float s) {
     uint16_t h;
@@ -68,10 +64,6 @@ static inline uint16_t to_brain(float s) {
         h = (u.i >> 16) | 64; /* force to quiet */
         return h;
     }
-    if (!(u.i & 0x7f800000)) { /* subnormal */
-        h = (u.i & 0x80000000) >> 16; /* flush to zero */
-        return h;
-    }
     return (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
 }