Make more ML improvements

- Fix UX issues with llama.com
- Do housekeeping on libm code
- Add more vectorization to GGML
- Get GGJT quantizer programs working well
- Have the quantizer keep the output layer as f16c
- Prefetching improves performance 15% if you use fewer threads

libc/tinymath/acoshl.c

@@ -38,6 +38,7 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/math.h"
 #include "libc/tinymath/freebsd.internal.h"
+#if !(LDBL_MANT_DIG == 53 && LDBL_MAX_EXP == 1024)
 
 asm(".ident\t\"\\n\\n\
 FreeBSD libm (BSD-2 License)\\n\
@@ -62,8 +63,6 @@ asm(".include \"libc/disclaimer.inc\"");
 #error "Unsupported long double format"
 #endif
 
-#define	BIAS	(LDBL_MAX_EXP - 1)
-
 static const double
 one	= 1.0;
 
@@ -108,3 +107,5 @@ acoshl(long double x)
 	    RETURNI(log1pl(t+sqrtl(2.0*t+t*t)));
 	}
 }
+
+#endif /* long double is long */