Yield the releasing goroutine's P to a blocked acquirer

[mikluko]

Fixes #46.

Release, Destroy and Hijack wake the first blocked acquirer by releasing the semaphore, which closes that waiter's channel. The runtime queues a goroutine woken that way as runnext of the waker's P, where it runs only once the waker blocks or another P steals it. A releaser that keeps the CPU after Release therefore leaves the resource unused for as long as it runs.

The fix is a runtime.Gosched() after the semaphore release, with the mutex already unlocked, taken only while a goroutine is blocked in Acquire; an atomic counter around acquireSem.Acquire says whether one is. The uncontended path is unchanged.

BenchmarkAcquire_ReleaseThenWork (first commit) is a pool of 4 under ~100 goroutines that hold for 1 ms and keep the CPU for 1 ms after releasing, -cpu 4 -count 3:

                          before                     after
linux/arm64, Go 1.26.7    411-425 µs/op             255-256 µs/op
darwin/arm64, Go 1.27.1   411-477 µs/op             254 µs/op

The same effect on the bare x/sync semaphore, and the per-acquire residue it corresponds to, is in the reproduction attached to #46: https://gist.github.com/mikluko/57dbb8a46e2d38b9b8b56cb4d7675f77

CreateResource and the asynchronous construction in initResourceValue release the semaphore too and are left alone: the constructing goroutine exits right after its release, and CreateResource is a warm-up call.

go vet reports pool_test.go:1308: call to (*testing.B).Fatal from a non-test goroutine on master already; not touched here.

[jackc]

This change seems to have a catastrophic performance impact when under heavy CPU load.

dev@lima-pgx-dev-vm ~/puddle (master) % go test -vet=off -run '^$' \
  -bench '^BenchmarkAcquire_ReleaseAfterAcquireWithCPULoad$' \
  -cpu 4 -benchtime=20x -count=1
goos: linux
goarch: arm64
pkg: github.com/jackc/puddle/v2
BenchmarkAcquire_ReleaseAfterAcquireWithCPULoad-4   	      20	      1525 ns/op
PASS
ok  	github.com/jackc/puddle/v2	2.103s
dev@lima-pgx-dev-vm ~/puddle (master) % cd ../puddle-pr47
dev@lima-pgx-dev-vm ~/puddle-pr47 (codex/pr-47) % go test -vet=off -run '^$' \
  -bench '^BenchmarkAcquire_ReleaseAfterAcquireWithCPULoad$' \
  -cpu 4 -benchtime=20x -count=1
goos: linux
goarch: arm64
pkg: github.com/jackc/puddle/v2
BenchmarkAcquire_ReleaseAfterAcquireWithCPULoad-4   	      20	 383744378 ns/op
PASS
ok  	github.com/jackc/puddle/v2	9.701s

[mikluko]

I don't see how to fix it without pulling patched version of semphore into the codebase. Releaser keeping the CPU is not an obvious edge case. Doesn't look like it warrants the scale of required changes. I think I will address it downstream. Sorry for the noise.