The original idea was to always lock `write_mutex` before acquiring an `InnerPool.semaphore` permit
to avoid ABBA deadlocks, but when refactoring a PR for b696a242fc,
that was forgotten.
This doesn't really change the program flow as we have `Context::housekeeping_mutex` anyway,
just simplifies the code.
This change is mainly to avoid exposing the write lock outside the pool module.
To avoid deadlocks, outside code should work only with the pooled connections
and use no more than one connection per thread.
When connection pool is organized as a stack,
it always returns most recently used connection.
Because each connection has its own page cache,
using the connection with fresh cache improves performance.
I commented out `oauth2::tests::test_oauth_from_mx`
because it requires network connection,
turned off the Wi-Fi and ran the tests.
Before the change, with a queue:
```
$ hyperfine "cargo test"
Benchmark 1: cargo test
Time (mean ± σ): 56.424 s ± 4.515 s [User: 183.181 s, System: 128.156 s]
Range (min … max): 52.123 s … 68.193 s 10 runs
```
With a stack:
```
$ hyperfine "cargo test"
Benchmark 1: cargo test
Time (mean ± σ): 29.887 s ± 1.377 s [User: 101.226 s, System: 45.573 s]
Range (min … max): 26.591 s … 31.010 s 10 runs
```
On version 1.107.1:
```
$ hyperfine "cargo test"
Benchmark 1: cargo test
Time (mean ± σ): 43.658 s ± 1.079 s [User: 202.582 s, System: 50.723 s]
Range (min … max): 41.531 s … 45.170 s 10 runs
```
