aarch64: Enable run-time detection of FEAT_LRCPC3/FEAT_LSE128 in load/store

[taiki-e]

• (new) If FEAT_LSE2 is not available at compile-time, we want to do run-time detection of FEAT_LSE2 (we are already doing this as it greatly improves performance: aarch64: Support run-time detection of FEAT_LSE2 #126), so we do run-time detection of FEAT_LRCPC3/FEAT_LSE128 at the same time.
• (as-is) If FEAT_LSE2 is available at compile-time, we don't do run-time detection of FEAT_LRCPC3/FEAT_LSE128 for load/store at this time, since FEAT_LRCPC3/FEAT_LSE128 is not yet available for most CPUs.
  (macOS that doesn't have any FEAT_LRCPC3/FEAT_LSE128-enabled CPUs as of M4 is only a platform that currently enables FEAT_LSE2 at compile-time by default.)
• (as-is) We don't do run-time detection of FEAT_LSE128 for swap/fetch_and/fetch_or at this time, because FEAT_LSE128 is not yet available for most CPUs, but since swpp/ldclrp/ldsetp is wait-free, it would make sense to do run-time detection in the future.

Also this documents the exact instruction selection per operation as comments.

Run-time detection of these features is usually supported on platforms that support run-time detection of FEAT_LSE2 and was implemented in this crate as a test-only implementation over 1 years ago (.e.g., f5daf34, e12b4bd)