riscv: Support 128-bit atomics for RV64 and 64-bit atomics for RV32 (Zacas extension)

[taiki-e]

This supports 128-bit atomics for riscv64 and 64-bit atomics for riscv32 with Zacas extension (ratified more than half a year ago) enabled.

Note:

• Currently Zacas extension support is marked as experimental in LLVM: so -C target-feature=+experimental-zacas instead of -C target-feature=+zacas is needed for now.
  • As the name indicates, support for this extension on our part at this time will also be somewhat experimental.
  • Perhaps we cannot merge this PR until LLVM marks this as non-experimental.
    • Now marked as experimental on our end too. And now only enabled for LLVM 19 because it is more experimental in LLVM 17/18 and "experimental-zacas" feature may no longer exist when it is marked as non-experimental in LLVM 20.
• The core part of runtime CPU feature detection has already been implemented for Linux/Android and tested for Linux on QEMU, but is test-only until zacas is marked as non-experimental by LLVM.
  • UPDATE: Supported as disabled by default for now in riscv: Support run-time detection of Zacas #183.
• Zacas extension provides DWCAS, but there is no DW load/store instruction in the ratified atomic-related extensions AFAIK, so a load/store implementation at this time would not be very efficient.

FYI @ibraheemdev