fix: validate final layer in VRAM stage compare

[booth-algo]

Summary

Make aten/vram_stage_compare.py validate the requested decoder layer instead of always reading layer 0 weights and O_full_0.

Changes:

• add optional layer_idx argument to compare_stages
• default layer_idx to the last W_o_<idx>.pt found in the build directory
• read W_o, W_gate, W_up, and W_down for that layer
• parse O_full_<layer_idx> from generated ASM comments
• expose optional CLI arg 3 for manual layer selection
• include layer_idx in the returned result dict

This supports multi-layer native emulator validation where the final-layer VRAM stage comparison is the authoritative check.

Validation

• py_compile passed for aten/vram_stage_compare.py
• Native 2-layer SmolVLM2 run validated final layer with inferred layer_idx=1
• O_proj + residual: 100.00% allclose, MSE 4.18e-06
• norm + FFN + residual + final_norm: 100.00% allclose, MSE 2.61e-05