Ease IMUL 5, h1 32 bits hash (~10 % for > 256 B).

Replace by a left shift and an addition. Generated code for 64 bits
version already map to the even most adequate `LEA`.

On a mid-2012 MBA,

benchmark                    old MB/s     new MB/s     speedup
Benchmark32_1-4              90.18        91.08        1.01x
Benchmark32_2-4              168.26       168.55       1.00x
Benchmark32_4-4              342.13       338.26       0.99x
Benchmark32_8-4              587.46       584.68       1.00x
Benchmark32_16-4             981.80       990.72       1.01x
Benchmark32_32-4             1349.67      1370.62      1.02x
Benchmark32_64-4             1684.94      1696.83      1.01x
Benchmark32_128-4            1748.24      1746.80      1.00x
Benchmark32_256-4            1805.96      1933.24      1.07x
Benchmark32_512-4            1845.56      2019.75      1.09x
Benchmark32_1024-4           1856.36      2031.61      1.09x
Benchmark32_2048-4           1890.43      2032.47      1.08x
Benchmark32_4096-4           1881.13      2085.36      1.11x
Benchmark32_8192-4           1889.83      2103.02      1.11x

murmur32.go

@@ -60,7 +60,7 @@ func (d *digest32) bmix(p []byte) (tail []byte) {
 
 		h1 ^= k1
 		h1 = (h1 << 13) | (h1 >> 19) // rotl32(h1, 13)
-		h1 = h1*5 + 0xe6546b64
+		h1 = h1*4 + h1 + 0xe6546b64
 	}
 	d.h1 = h1
 	return p[nblocks*d.Size():]
@@ -134,7 +134,7 @@ func Sum32WithSeed(data []byte, seed uint32) uint32 {
 
 		h1 ^= k1
 		h1 = (h1 << 13) | (h1 >> 19) // rotl32(h1, 13)
-		h1 = h1*5 + 0xe6546b64
+		h1 = h1*4 + h1 + 0xe6546b64
 	}
 
 	tail := data[nblocks*4:]