[CI] Make wrapper2/wrapper3 exercise small systolic-array configs

[YWHyuk]

Problem

The test-pytorchsim-wrapper2 CI job injects vpu_num_lanes=32 and vpu_spad_size_kb_per_lane=32 as docker env vars to test a small configuration. But since the unified-config refactor (9db0f2ce), extension_config.py reads these values only from the TOGSim YAML, never from the environment:

if name == "vpu_num_lanes":
    return config_yaml["vpu_num_lanes"]   # os.environ is never consulted

So the env vars are silently ignored and wrapper2 falls back to the default 128x128 config, making it an exact duplicate of wrapper1.

TOGSim (the C++ cycle sim) also reads vpu config from the --config YAML, not from env, so reviving an env override would only desync codegen from the cycle model and fight the unified-config design.

Fix (single source of truth)

Select a config YAML via TOGSIM_CONFIG so one file drives both codegen and the TOGSim cycle model. vpu_num_lanes also sets the systolic-array dimension, so each config exercises a different array size.

• configs: add systolic_ws_32x32_c1_simple_noc_tpuv3.yml (32x32 array, 32 KB/lane SPAD) and systolic_ws_8x8_c1_simple_noc_tpuv3.yml (8x8 array, 32 KB/lane SPAD); both identical to the tpuv3 default otherwise.
• pytorchsim_test.yml: replace the vector_lane/spad_size number inputs with togsim_config (string) and run_accuracy (bool). Every job now sets -e TOGSIM_CONFIG instead of the dead vpu_* env vars. The accuracy/speedup job is gated on run_accuracy instead of vector_lane == 128.
• docker-image.yml: wrapper1 -> 128x128 config + run_accuracy: true; wrapper2 -> 32x32 config; wrapper3 -> 8x8 config (run_accuracy: false).

Verification

• Both workflow YAMLs parse cleanly.
• With TOGSIM_CONFIG pointed at each new config, extension_config reads vpu_num_lanes = 32 / 8 and SPAD = 32 KB/lane respectively; the default still reads 128/128. wrapper2 and wrapper3 now genuinely exercise the smaller array paths for codegen and the cycle sim consistently.