Fix simx flush arbiter

[ATWeatherly]

some bug fixes that arose when testing VM work

[tinebp]

Reviewed and accepted — thank you, this is a real bug and the diagnosis was correct.

Verification

Reproduced on master before the fix. The shape that breaks is any request count that is not a
multiple of the output count with more than one output:

./ci/blackbox.sh --driver=simx --app=demo --cores=3 \
  --perf=0 CONFIGS="-DVX_CFG_SOCKET_SIZE=3 -DVX_CFG_NUM_ICACHES=2"

3 cores over 2 shared icaches → log2ceil(3/2) = log2ceil(1) = 0, so each output arbitrates
over a single input and the third core's fetch port is never polled. The run hangs at boot
(timeout, rc=124). With log2ceil(CDIV(3,2)) = 1 it passes. VX_stream_arb has always sized its
groups by ceiling division, so this was a straight SimX/RTL model divergence.

Why CI missed it

Every catalog case that shares an icache used a power-of-two core:cache ratio, so
inputs % outputs == 0 held everywhere and the floor/ceiling distinction never showed. I added
one case in the cache category that pins the non-power-of-two shape:

- id: demo-shared-icache-odd
  via: blackbox
  drivers: [simx]
  app: demo
  configs: -DVX_CFG_SOCKET_SIZE=3 -DVX_CFG_NUM_ICACHES=2
  shape: { cores: 3 }

Validated both ways: it hangs under floor grouping and passes under ceiling grouping.

What landed

a3e7bf222 on master, authored to you, with the arbiter fix, the catalog case, and a
CHANGELOG entry. The full tests/regression suite and the kernel conform suite stay green on
simx and rtlsim.

The other half of this PR — the cache-flush walk gating on size() rather than TFifo::empty() —
had already landed independently via #394, which is why this branch conflicted with master and
could not be updated in place. Only the arbiter half was rebased forward. Closing here since the
work is merged; the commit carries your authorship.