fix(qualcomm): harden compile_npu.py with CLI options and container format pre-flight checks (Fixes #121)

[jhaabhijeet864]

Summary

Resolves #121 by replacing hardcoded machine paths in compile_npu.py with standard CLI arguments, environment variable detection, and adding container format pre-flight inspection.

Key Changes

1. CLI & Environment Hardening:
   • Replaced hardcoded author paths with dynamic argparse flags (--model_path, --output_dir, --soc_model, --qairt_root) and fallback to $QAIRT_ROOT.
   • Added support for multiple Snapdragon SoC targets (SM8750 / Snapdragon 8 Elite, SM8650 / 8 Gen 3, SM8550 / 8 Gen 2, etc.).
2. Container Format Pre-Flight Diagnostics:
   • Added magic-byte inspection (check_model_magic_bytes): informs users when a .litertlm container (RTLM) is passed into aot_compile (which requires flat .tflite TFL3 graphs) with actionable guidance, preventing low-level C++ FlatBuffer crashes.
3. Documentation Cleanup:
   • Cleaned up NPU_COMPILATION_GUIDE.md across llm_chatbot_npu and gemma3/npu to document the new CLI parameters.

Verification

• Tested compile_npu.py --help CLI parameter parsing.
• Tested magic-byte pre-flight checks against dummy RTLM and TFL3 models.

Fixes #121