aarch64: Support run-time detection of FEAT_LSE2

[taiki-e]

This supports run-time detection of FEAT_LSE2 in 128-bit atomic load/store.

This will greatly improve performance of 128-bit atomic load/store on hardware such as graviton 3, apple m1[^1] (even without -C target-feature=+lse2). In aarch64 macOS, FEAT_LSE2 is enabled at compile time, so it is already optimized and run-time detection is unnecessary and unused. However, Linux and other systems running on m1 will benefit from run-time detection.

The following is the result of the microbenchmark on aarch64 Linux on apple m1.

bench_portable_atomic_arch/u128_load
                        time:   [1.5079 ns 1.5090 ns 1.5104 ns]
                        change: [-79.046% -78.953% -78.891%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 12 outliers among 100 measurements (12.00%)
  5 (5.00%) high mild
  7 (7.00%) high severe
bench_portable_atomic_arch/u128_store
                        time:   [1.1308 ns 1.1317 ns 1.1326 ns]
                        change: [-80.159% -80.046% -79.970%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 15 outliers among 100 measurements (15.00%)
  8 (8.00%) high mild
  7 (7.00%) high severe
bench_portable_atomic_arch/u128_concurrent_load
                        time:   [78.471 µs 79.294 µs 80.459 µs]
                        change: [-51.261% -50.628% -49.878%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
  2 (2.00%) high mild
  4 (4.00%) high severe
bench_portable_atomic_arch/u128_concurrent_load_store
                        time:   [125.96 µs 126.72 µs 127.65 µs]
                        change: [-62.113% -61.410% -60.740%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 7 outliers among 100 measurements (7.00%)
  5 (5.00%) high mild
  2 (2.00%) high severe
bench_portable_atomic_arch/u128_concurrent_store
                        time:   [136.19 µs 136.87 µs 137.58 µs]
                        change: [-53.565% -52.888% -52.263%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 7 outliers among 100 measurements (7.00%)
  3 (3.00%) high mild
  4 (4.00%) high severe

The detection implementation uses an implementation that has already existed for testing.

cc #10