strconv: use the Eisel-Lemire ParseFloat algorithm

strconv: use the Eisel-Lemire ParseFloat algorithm

Also fix BenchmarkAtof64Random* to initialize the test data when none
of the TestAtof* tests are run.

Passing "go test -test.count=5 -test.run=xxx -test.bench=Atof64" on to
benchstat:

name                  old time/op  new time/op  delta
Atof64Decimal-4       47.9ns ± 0%  48.3ns ± 1%     ~     (p=0.238 n=4+5)
Atof64Float-4         58.3ns ± 3%  57.7ns ± 0%     ~     (p=0.151 n=5+5)
Atof64FloatExp-4       107ns ± 0%    71ns ± 1%  -33.89%  (p=0.016 n=4+5)
Atof64Big-4            163ns ± 0%   166ns ± 2%     ~     (p=0.159 n=4+5)
Atof64RandomBits-4     299ns ± 1%   166ns ± 1%  -44.41%  (p=0.008 n=5+5)
Atof64RandomFloats-4   188ns ± 1%   144ns ± 0%  -23.03%  (p=0.008 n=5+5)

The canada.json file from github.com/miloyip/nativejson-benchmark is
full of geospatial coordinates (i.e. numbers). With this program:

    src, _ := ioutil.ReadFile("canada.json")
    for i := 0; i < 5; i++ {
        now := time.Now()
        for j := 0; j < 10; j++ {
            dst := interface{}(nil)
            if err := json.Unmarshal(src, &dst); err != nil {
                log.Fatal(err)
            }
        }
        fmt.Println(time.Since(now))
    }

Median of the 5 printed numbers, lower is better.
Before: 760.819549ms
After:  702.651646ms
Ratio:  1.08x

The new detailedPowersOfTen table weighs in at 596 * 16 = 9536 bytes,
but some of that weight gain can be clawed back, in a follow-up commit,
that folds in the existing powersOfTen table in extfloat.go.

RELNOTE=yes

Change-Id: I3953110deaa1f5f6941e88e8417c4665b649ed80
Reviewed-on: https://go-review.googlesource.com/c/go/+/260858
Run-TryBot: Nigel Tao <nigeltao@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
Trust: Nigel Tao <nigeltao@golang.org>