Minor modifications to the optimized amd64 implememntation.

* Reduce window size: reduces size of the lookup tables by 40%
* Revised scalar inversion formula, with less operations
* Field square function now uses intental loop, saving call overhead

This change will serve as a basis for an arm64 implementation.

Performance results on Skylake MacBook Pro:

pkg:crypto/elliptic goos:darwin goarch:amd64
BaseMultP256      17.8µs ± 1%    17.5µs ± 1%  -1.41%  (p=0.003 n=10+10)
ScalarMultP256    70.7µs ± 1%    68.9µs ± 2%  -2.57%  (p=0.000 n=9+9)
pkg:crypto/ecdsa goos:darwin goarch:amd64
SignP256          32.7µs ± 1%    31.4µs ± 1%  -3.96%  (p=0.000 n=10+8)
VerifyP256        95.1µs ± 1%    93.5µs ± 2%  -1.73%  (p=0.001 n=10+9)

name            old alloc/op   new alloc/op   delta
pkg:crypto/elliptic goos:darwin goarch:amd64
BaseMultP256        288B ± 0%      288B ± 0%    ~     (all equal)
ScalarMultP256      256B ± 0%      256B ± 0%    ~     (all equal)
pkg:crypto/ecdsa goos:darwin goarch:amd64
SignP256          2.90kB ± 0%    2.90kB ± 0%    ~     (all equal)
VerifyP256          976B ± 0%      976B ± 0%    ~     (all equal)

name            old allocs/op  new allocs/op  delta
pkg:crypto/elliptic goos:darwin goarch:amd64
BaseMultP256        6.00 ± 0%      6.00 ± 0%    ~     (all equal)
ScalarMultP256      5.00 ± 0%      5.00 ± 0%    ~     (all equal)
pkg:crypto/ecdsa goos:darwin goarch:amd64
SignP256            34.0 ± 0%      34.0 ± 0%    ~     (all equal)
VerifyP256          17.0 ± 0%      17.0 ± 0%    ~     (all equal)

Change-Id: I3f0e2e197a54e7bc7916dedc5dbf085e2c4aea24
Reviewed-on: https://go-review.googlesource.com/99622Reviewed-by: Vlad Krasnov <vlad@cloudflare.com>
Reviewed-by: Filippo Valsorda <filippo@golang.org>
Run-TryBot: Vlad Krasnov <vlad@cloudflare.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>