*: remove goroutine pool

[tiancaiamao]

What problem does this PR solve?

Benchmark on Go1.10:

➜ goroutine_pool git:(master) ✗ go test -test.bench Benchmark
goos: linux
goarch: amd64
pkg: github.com/pingcap/tidb/util/goroutine_pool
BenchmarkGoPool-4 5000000 382 ns/op
BenchmarkGo-4 10000000 163 ns/op
BenchmarkMorestackPool-4 1000000 1092 ns/op
BenchmarkMoreStack-4 500000 3012 ns/op
PASS
ok github.com/pingcap/tidb/util/goroutine_pool 8.303s

Benchmark on Go1.11:

➜ goroutine_pool git:(master) ✗ go test -test.bench Benchmark
goos: linux
goarch: amd64
pkg: github.com/pingcap/tidb/util/goroutine_pool
BenchmarkGoPool-4 5000000 326 ns/op
BenchmarkGo-4 10000000 176 ns/op
BenchmarkMorestackPool-4 2000000 719 ns/op
BenchmarkMoreStack-4 5000000 391 ns/op
PASS
ok github.com/pingcap/tidb/util/goroutine_pool 10.035s

We can see that BenchmarkMoreStack has improved dramatically, it's even faster than a pooled version.
So it's time to remove the goroutine pool.

What is changed and how it works?

goroutine pool was introduced to handle stack copy cost #3752 #3753

Go1.11 has many optimizations for stack copy:

https://go-review.googlesource.com/c/go/+/94029 runtime: speed up stack copying a little
https://go-review.googlesource.com/c/go/+/104737 runtime: avoid calling adjustpointers unnecessarily
https://go-review.googlesource.com/c/go/+/108945 runtime: add fast version of getArgInfo
https://go-review.googlesource.com/c/go/+/109716 runtime: iterate over set bits in adjustpointers
https://go-review.googlesource.com/c/go/+/109001 runtime: allow inlining of stackmapdata
https://go-review.googlesource.com/c/go/+/104175 cmd/compile: shrink liveness maps
These optimizations collectively improve stack copy performance by ~2X

After upgrading to Go1.1, goroutine pool is not necessary any more.

Check List

Tests

• No code

Code changes

remove goroutine_pool package and update the code

[ngaut]

LGTM

[winkyao]

Cool.

[winkyao]

LGTM

[tiancaiamao]

/run-all-tests