net: TestDialParallel failures due to longer-than-expected delay on Windows

[bcmills]

--- FAIL: TestDialParallel (6.16s)
    dial_test.go:173: dialClosedPort: measured delay 1.0028044s
    dial_test.go:313: #5: got 1.2757236s; want <= 1.20336528s
FAIL
FAIL	net	55.514s

greplogs --dashboard -md -l -e 'FAIL: TestDialParallel .*(?:\n .*)*got .*; want <= .*'

2022-04-05T18:01:26-9e16cc1/windows-amd64-longtest

Note that the test currently uses a hard-coded heuristic for the timing bounds:
https://cs.opensource.google/go/go/+/master:src/net/dial_test.go;l=302-309;drc=da7891f6f36c48f2931ed916ed305330c06f9bd7

(See previously #35616; CC @ianlancetaylor @neild.)

[bcmills]

Since windows/amd64 is a first-class port, this is a release-blocker for Go 1.19.

(It may be possible to resolve by making the timing bounds even more lax, or by somehow eliminating the timing bounds entirely if the test is meaningful without them.)

[prattmic]

FWIW, I've been attempting to reproduce a different Windows issue, but have managed to reproduce this twice thus far in ~1hr of testing with 25 windows-amd64-longtest gomotes running all.bat using https://github.com/mknyszek/goswarm/.

Edit: I got 58(!) cases of this over the weekend.

[gopherbot]

Change https://go.dev/cl/408354 mentions this issue: net: add even more timing slop for TestDialParallel