In asm, load primes from globals rather than using macros

This also improves the speed slightly. Maybe just because of smaller
code size?

benchmark                           old ns/op     new ns/op     delta
BenchmarkHashes/xxhash,n=5b-4       9.69          9.23          -4.75%
BenchmarkHashes/xxhash,n=20b-4      11.7          11.4          -2.56%
BenchmarkHashes/xxhash,n=100b-4     24.4          23.6          -3.28%
BenchmarkHashes/xxhash,n=4KB-4      384           375           -2.34%
BenchmarkHashes/xxhash,n=10MB-4     1113137       1089693       -2.11%

benchmark                           old MB/s     new MB/s     speedup
BenchmarkHashes/xxhash,n=5b-4       515.96       541.56       1.05x
BenchmarkHashes/xxhash,n=20b-4      1704.45      1761.10      1.03x
BenchmarkHashes/xxhash,n=100b-4     4092.55      4245.48      1.04x
BenchmarkHashes/xxhash,n=4KB-4      10406.76     10638.45     1.02x
BenchmarkHashes/xxhash,n=10MB-4     8983.62      9176.89      1.02x