50% fewer mallocs in HTML tokenization, resulting in 25% fewer mallocs
in parsing go1.html.

Making the parser use integer comparisons instead of string comparisons
will be a follow-up CL, to be co-ordinated with Andy Balholm's work.

exp/html benchmarks before/after:

BenchmarkParser	     500	   4754294 ns/op	  16.44 MB/s
        parse_test.go:409: 500 iterations, 14651 mallocs per iteration
BenchmarkRawLevelTokenizer	    2000	    903481 ns/op	  86.51 MB/s
        token_test.go:678: 2000 iterations, 28 mallocs per iteration
BenchmarkLowLevelTokenizer	    2000	   1260485 ns/op	  62.01 MB/s
        token_test.go:678: 2000 iterations, 41 mallocs per iteration
BenchmarkHighLevelTokenizer	    1000	   2165964 ns/op	  36.09 MB/s
        token_test.go:678: 1000 iterations, 6616 mallocs per iteration

BenchmarkParser	     500	   4664912 ns/op	  16.76 MB/s
        parse_test.go:409: 500 iterations, 11266 mallocs per iteration
BenchmarkRawLevelTokenizer	    2000	    903065 ns/op	  86.55 MB/s
        token_test.go:678: 2000 iterations, 28 mallocs per iteration
BenchmarkLowLevelTokenizer	    2000	   1260032 ns/op	  62.03 MB/s
        token_test.go:678: 2000 iterations, 41 mallocs per iteration
BenchmarkHighLevelTokenizer	    1000	   2143356 ns/op	  36.47 MB/s
        token_test.go:678: 1000 iterations, 3231 mallocs per iteration

R=r, rsc, rogpeppe
CC=andybalholm, golang-dev
https://golang.org/cl/6255062