Increase ExecuTorch thread stack size

Sync

…-Executorch into whole-archive

This reverts commit 81c34eb.

- Install tosa-tools==0.0.4 from Test PyPI using --extra-index-url (pre-built wheel)
- Avoids ~3 minute C++ source compilation of reference_model and serialization libs
- Patch setup.sh to skip tosa-tools source builds via sed (git clone still runs)
- Reduces total Docker build time by ~3 minutes (from ~32min to ~29min)
- Uses --extra-index-url to resolve numpy from main PyPI while fetching tosa-tools from Test PyPI

- Replace tosa-tools==0.0.4 from Test PyPI with tosa-tools==2026.2.0 from official PyPI
- Remove --extra-index-url workaround (no longer needed)
- Simplify installation to use main PyPI directly

Add scripts to translate ExecuTorch operator names (aten::, quantized_decomposed::)
into PyTorch::ExecuTorch CMSIS-Pack component references, enabling AI layer
generation without Docker builds.

New files:
- scripts/generate_pack_clayer.py: Python script with operator-to-component mapping
  for 143 portable + 9 quantized operators. Supports input from .pte model files,
  operators list files, selected_operators.yaml, or command-line lists.
- scripts/generate_pack_layer.sh: Shell wrapper combining model conversion and
  pack-based layer generation.
- documentation/PACK_BASED_LAYER.md: Comprehensive documentation including mapping
  tables, usage examples, and comparison with the Docker-based workflow.

Modified:
- .vscode/tasks.json: Added 'Pack: Generate AI Layer from Model' and
  'Pack: Generate AI Layer from Operators File' VS Code tasks.

Replace the 7-step source-compilation workflow (stage1/stage2 CMake builds,
source layer generation, header patching, artifact packaging) with a 3-step
pack-based workflow:
  1. Convert PyTorch model to .pte (aot_model.py)
  2. Convert .pte to C header (pte_to_header.py)
  3. Generate pack-based ai_layer.clayer.yml (generate_pack_clayer.py)

The generated layer references PyTorch::ExecuTorch pack components instead of
compiled source files, eliminating the need for ExecuTorch source compilation.

Extend the model from a simple Add (fully delegated to Ethos-U) to
AddWithPostProcessing which includes CPU-side portable operators:
  - view_copy, mul, add, sigmoid, unsqueeze_copy, softmax

These ops stay on the CPU (not delegated to NPU) so they exercise the
pack-based operator component selection in the generated clayer.yml.

The previous model used set_global() which quantized ALL ops including
the post-processing chain (view, mul, sigmoid, softmax). The EthosU
partitioner then delegated everything to the NPU, leaving zero CPU ops.

Fix: split the add into an InnerAdd submodule and use
quantizer.set_module_name('inner_add', config) so only the add gets
quantized and delegated. The post-processing ops stay as float and
will appear in the .pte as portable CPU operators.

Parses model_conversion (Vela) and generate_pack_layer logs to produce
a structured report with selected operators table, TOSA graphs, NPU
performance summary, network summary, and final exported program graph.

Integrated as Step 4 in local_workflow.sh (runs after pack layer gen).