Reliability improvements for `RawMutex`

[Radonoxius]

I have done some modifications to lock_slow and unlock of RawMutex.

These are the benchmark results currently (on 8-core Aarch64 device):

- Running with 8 threads
- 5 iterations inside lock, 5 iterations outside lock
- 5 seconds per test
        name         |    average     |     median     |    std.dev.   
parking_lot::Mutex   |   1036.994 kHz |   1669.948 kHz |    851.467 kHz
std::sync::Mutex     |    677.364 kHz |    680.531 kHz |     12.590 kHz
pthread_mutex_t      |    633.672 kHz |   1008.571 kHz |    378.294 kHz

After the modifications, its the following:

- Running with 8 threads
- 5 iterations inside lock, 5 iterations outside lock
- 5 seconds per test
        name         |    average     |     median     |    std.dev.   
parking_lot::Mutex   |   1399.458 kHz |   1440.332 kHz |     69.834 kHz
std::sync::Mutex     |    670.454 kHz |    672.427 kHz |     12.889 kHz
pthread_mutex_t      |    638.275 kHz |    900.675 kHz |    348.836 kHz

Would you be interested in a PR for these modifications?

The optimisation involves the removal of unnecessary CAS (weak) operations. It might fare even better on x86_64.

[Radonoxius]

Would you be interested in a PR for these modifications?