Add experimental aarch64-pc-msys support

[Kreijstal]

This is an early port of msys2-runtime to native Windows ARM64
(aarch64-pc-msys). It builds on the AArch64 Cygwin work in
Windows-on-ARM-Experiments,
adapts it to the MSYS2 3.6.9 tree, and adds a Clang-based cross-build path.

I tested the resulting runtime and Bash on Windows 11 ARM64 in QEMU/HVF. I am
opening this as a draft because the basic process model works, but several
runtime issues and the source-attribution cleanup still need to be addressed.

Changes

• recognize aarch64-pc-msys in the build system
• add the AArch64 newlib and PE/COFF support needed by the runtime
• adapt winsup startup, TLS, imports, exceptions, signals, pthreads, memory
  allocation, and fork handling for ARM64
• allow the target to be cross-built with Clang and ARM64 PE binutils
• document the toolchain and reproduction procedure

Testing

All existing GitHub Actions jobs pass, including the GCC, Clang, MSYS, and
CLANGARM64 configurations.

I also rebuilt commit 8e82becb50b8bf02fa0d0dd9cd6d53671d3a1347
from a clean checkout with the pinned LLVM, binutils, mingw-w64, and
LLVM-MinGW inputs. The prepared DLL was a native AArch64 PE32+ image with
SHA-256:

e69dd757dab8ca0148dbf0f0e95278afab70e56f527b1406bf9eb1d86ba653ed  msys-2.0.dll

On Windows 11 ARM64 under QEMU/HVF, that exact DLL passed:

• native LoadLibrary and dll_crt0 resolution
• runtime startup and external-process execution
• native fork and pipe/fork probes, 10/10 runs each
• Bash 5.3 startup and smoke tests, 5/5 runs
• command substitution and subshell exit-status checks, 6/6 runs
• a focused sigsetjmp/siglongjmp probe, 5/5 runs; saved signal masks were
  restored and the savemask=0 case left the current mask unchanged

One earlier Bash status invocation ended with 0xC0000005. During the ten-run
fork sweep, one child copy attempt reported Win32 error 299 and then succeeded
through the runtime retry path. These observations are consistent with the
transient QEMU child-start behavior described below, so this is functional
validation rather than a stability claim.

The reproduction gist
contains the pinned inputs, build scripts, hashes, probes, and captured output.

Known issues

• long Bash argument lists can be corrupted
• /tmp setup currently emits a warning
• transient fork-child startup faults are retried up to four times
• a rare process fault can occur in QEMU before the runtime is loaded
• newly linked ARM64 executables still require the documented CRT post-link fix
• the full native ARM64 runtime testsuite is not yet running as a required CI
  job

Prior work

The runtime implementation is derived from the Windows-on-ARM-Experiments
effort rather than being a separate ARM64 port. Before this is ready for
review, I need to compare the runtime commits against that history, retain
attribution for the code and designs carried over, and remove changes that
duplicate newer WOA work. The QEMU reproduction and Clang/MSYS2 adaptation are
the parts specifically validated here.