examples/qwen3: add SwiGLU MLP component with equivalence tests and microbenchmark

[chhwang]

examples/qwen3: add SwiGLU MLP component with equivalence tests and microbenchmark

Q5 of the Qwen3-on-ARK component build-out. Three new files, no existing
files modified.

ark_mlp.py — ark_swiglu_mlp() performs gate/up/down projections via
torch.matmul and the SiLU·gate activation via F.silu(gate) * up.
ARK-native ark_silu_gate (composed ark.sigmoid + ark.mul) is retained
dormant: the upstream composed-graph planner bug that crashed Q4 at 4-D
shapes also crashes at the MLP intermediate shape (2048, 12288). Output
wrapped in ark.copy for .eval() API consistency.

test_mlp.py — 7 pytest cases gated with skipif(not cuda):
test_silu_gate (intermediate activation), test_mlp_small (B=1, S=16),
test_mlp_prefill (B=1, S=128), test_mlp_decode (B=1, S=1),
test_mlp_batch (B=2, S=16), test_mlp_output_shape, and
test_ark_silu_gate_large_xfail (subprocess spike at (2048, 12288)
documenting the upstream bug). Tolerance: atol/rtol=5e-3.

bench_mlp.py — microbenchmark comparing torch eager SwiGLUMLP vs
ark_swiglu_mlp at Qwen3-8B shapes (S=2048 prefill, S=1 decode). Banner
labels output "torch-only (ARK silu·gate deferred to upstream fix / Q10)."

Follows Q4 patterns: torch fallback for crashed ARK ops, small-config tests,
subprocess-isolated xfail for bug documentation.