Coroutine: Assembly entry on every CPU, tested with QEMU. v1.9.3

[winlinvip]

This PR starts every new coroutine in an assembly entry, _st_md_thread_start, on every Linux and macOS CPU, the way Windows x64 already does since v1.9.2. It also adds a QEMU toolchain that builds and tests ST for 10 Linux CPUs on one machine.

./auto/qemu.sh all            # every Linux CPU at once, with a PASS/FAIL summary
./auto/qemu.sh riscv64 utest  # one CPU, one run
./auto/darwin.sh x86_64       # macOS x86_64 under Rosetta 2

Coroutine entry

Before, st_thread_create saved the context, and a new coroutine started by returning a second time from that save, on a new stack. That is undefined behavior: the compiler may still use the creator's frame or registers after the save, the coroutine inherits the creator's callee-saved registers, and backtraces and unwinders have no clean bottom frame. It worked only because the code after the save is trivial.

• st_thread_create still saves the context, but only as a template. MD_INIT_THREAD_ENTRY in md.h then sets the saved SP, PC and frame pointer, so the first switch to the coroutine enters _st_md_thread_start on its own stack, which calls _st_thread_main.
• The entry is in md_linux.S, md_linux2.S and md_darwin.S for Linux x86_64, i386, aarch64, arm, riscv64, loongarch64, mips, mipsel, mips64 and mips64el, and macOS arm64 and x86_64. Each one aligns the stack as the ABI expects, nulls the frame pointer and return address, and marks the return address undefined in the CFI (and EHABI on arm), so a backtrace ends at the entry.
• Cygwin64 alone keeps the old start, under #ifndef MD_INIT_THREAD_ENTRY. It is not tested.
• docs/coroutine_entry.md describes the entry on every CPU, with a per-CPU table. docs/win64_coroutine.md keeps the Windows x64 details.

Context save fixes found by QEMU

• riscv64 saves and restores fs0–fs11, MIPS the callee-saved FP registers (o32 and n64), and loongarch64 $f24–$f31. Before, a coroutine switch lost them. Soft-float builds save none.
• The arm jmpbuf grows to fit what the save stores (VFP and iWMMXt). Before, the save wrote past the jmpbuf. md.h now fails to compile if a jmpbuf is too small, and a utest checks on every CPU that the save writes nothing past it.
• The MIPS JB_* defines are fixed, and the porting tool builds on big-endian MIPS.
• The thread tool uses a 16 KB frame on the default stack, so it passes on macOS x86_64, where ___chkstk_darwin failed on the larger one.

Tests

• utest/st_utest_entry.cpp checks the saved PC and frame pointer accessors (MD_GET_PC, MD_GET_FP) on every CPU, and that a new coroutine starts in the entry.
• tools/backtrace checks that the outermost frame of a backtrace in a coroutine is inside _st_md_thread_start, on every platform but Cygwin64. On Windows it follows the MSVC incremental-link thunk.
• The Linux library and tools/backtrace build with -funwind-tables, so the backtrace works on arm and MIPS.
• The riscv64 FIFO write timeout test starts its timer after building the stream, so it passes under QEMU.

QEMU and Rosetta toolchain

• auto/qemu.sh <cpu> [run] cross-builds ST with g++-<cpu> in one ubuntu:24.04 container and runs the utest and tools with qemu-user 10.2. x86_64 runs natively in an amd64 container instead. Runs are utest, tools, tools-malloc, asan (x86_64 and aarch64) and all.
• auto/qemu.sh all runs every CPU in its own container, ST_QEMU_JOBS at once (5 by default), and prints a per-CPU table. All 10 CPUs pass in about 90 s.
• The image is tagged by a hash of auto/qemu/, so a Dockerfile change builds a new image. After a build, the script removes the older st-qemu tags.
• auto/qemu.sh <cpu> shell opens a shell in the image, with gdb-multiarch to debug a CPU through qemu-user's gdb stub.
• TARGETDIR builds every output in its own folder, so several CPUs can build in one checkout at once. Default builds are unchanged.
• auto/darwin.sh arm64|x86_64 builds and runs ST on macOS, x86_64 under Rosetta 2.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.01%. Comparing base (099b714) to head (96863b3).

Additional details and impacted files

@@           Coverage Diff           @@
##              srs      #42   +/-   ##
=======================================
  Coverage   96.01%   96.01%           
=======================================
  Files           8        8           
  Lines        1129     1129           
  Branches      287      286    -1     
=======================================
  Hits         1084     1084           
  Partials       45       45           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.