feat(sim): add ring-buffer wrapper for gvsoc trace to bound disk usage

[runwangdl]

When debugging gvsoc hangs on Siracusa (which has no working UART) the only progress signal is gvsoc's --trace output. A full --trace=insn run on a real-size network produces tens of GB before the simulation completes — enough to fill the build host's disk on a single attempt.

What this PR adds

scripts/ring_tee.py (~170 lines, pure stdlib)

Reads stdin, rotates writes across N files of fixed size each, and discards the oldest content once keep * size is full. Three runtime features make it useful for hang debugging:

• 5-second stderr heartbeat — distinguishes "gvsoc itself froze" (heartbeat stops because stdin went silent) from "simulated firmware deadlocked" (heartbeat continues but simulated PC stays still in the trace).
• SIGUSR1 handler — writes <prefix>.snapshot as a chronological concatenation of every file still on disk (oldest → newest), so a single kill -USR1 <pid> grabs the recent window at the moment of the hang.
• Ring rotation — total disk use is bounded by keep * size (default 600 MB) regardless of how long the simulation runs.

cmake/simulation.cmake options

Option	Default	Purpose
GVSOC_RING_TRACE	OFF	Enable the wrapper. Default OFF — behaviour identical to before when not set.
GVSOC_RING_TRACE_SIZE	500M	Per-file size, K/M/G suffix accepted.
GVSOC_RING_TRACE_KEEP	3	Number of rotating files.

When GVSOC_RING_TRACE=ON, the gvsoc_<name> target pipes gvsoc's stderr through ring_tee.py via bash process substitution.

Usage

cmake -DGVSOC_RING_TRACE=ON \
      -DGVSOC_RING_TRACE_SIZE=500M \
      -DGVSOC_RING_TRACE_KEEP=3 \
      ...
make gvsoc_resnet8_train

Trace files land at:

• <build>/gvsoc_workdir/gv_trace.0 / .1 / .2
• <build>/gvsoc_workdir/gv_trace.current — symlink → currently-writing file

When stuck:

kill -USR1 $(pgrep -f ring_tee.py)
# inspect <build>/gvsoc_workdir/gv_trace.snapshot (chronological full window)

Known limitations (follow-up)

1. This gvsoc build emits --trace=insn output on stdout, not stderr — to capture it via this wrapper you currently need a hand-written bash script that pipes stdout into ring_tee, rather than using the CMake target.
2. The CMake macro's bash -c '...' quoting loses process substitution when expanded by make's /bin/sh (POSIX mode >(...) is treated as a literal filename and aborts).

Both limitations are left to a follow-up; this PR lands the wrapper script and CMake plumbing so the infrastructure is in place. A bash wrapper script using ring_tee directly (bypassing the CMake macro) was used successfully today to debug the PR #19 graph-I/O promote hang.