crypto/tls: adjust dynamic record sizes to grow arithmetically

crypto/tls: adjust dynamic record sizes to grow arithmetically

The current code, introduced after Go 1.6 to improve latency on
low-bandwidth connections, sends 1 kB packets until 1 MB has been sent,
and then sends 16 kB packets (the maximum record size).

Unfortunately this decreases throughput for 1-16 MB responses by 20% or so.

Following discussion on #15713, change cutoff to 128 kB sent
and also grow the size allowed for successive packets:
1 kB, 2 kB, 3 kB, ..., 15 kB, 16 kB.
This fixes the throughput problems: the overhead is now closer to 2%.

I hope this still helps with latency but I don't have a great way to test it.
At the least, it's not worse than Go 1.6.

Comparing MaxPacket vs DynamicPacket benchmarks:

name              maxpkt time/op  dyn. time/op delta
Throughput/1MB-8    5.07ms ± 7%   5.21ms ± 7%  +2.73%  (p=0.023 n=16+16)
Throughput/2MB-8   15.7ms ±201%    8.4ms ± 5%    ~     (p=0.604 n=20+16)
Throughput/4MB-8    14.3ms ± 1%   14.5ms ± 1%  +1.53%  (p=0.000 n=16+16)
Throughput/8MB-8    26.6ms ± 1%   26.8ms ± 1%  +0.47%  (p=0.003 n=19+18)
Throughput/16MB-8   51.0ms ± 1%   51.3ms ± 1%  +0.47%  (p=0.000 n=20+20)
Throughput/32MB-8    100ms ± 1%    100ms ± 1%  +0.24%  (p=0.033 n=20+20)
Throughput/64MB-8    197ms ± 0%    198ms ± 0%  +0.56%   (p=0.000 n=18+7)

The small MB runs are bimodal in both cases, probably GC pauses.
But there's clearly no general slowdown anymore.

Fixes #15713.

Change-Id: I5fc44680ba71812d24baac142bceee0e23f2e382
Reviewed-on: https://go-review.googlesource.com/23487
Reviewed-by: Ian Lance Taylor <iant@golang.org>