FEAT: Add DeepONet surrogate method (proof of concept)

[nikolasborrel]

Why

• Adds DeepONet as a fourth method — a neural-operator surrogate predicting impulse responses instead of solving the wave equation.
• develop has an unregistered deeponet_method/ prototype; this replaces it with a packaged module and wires it into methods-config.json.
• Unlike the other solvers, it is a two-stage pipeline: DG runs first and writes its field data, that data trains DeepONet, then inference produces the IRs. One run = generate → train → infer.

Changes

Area	Change
deeponet_interface/ (new package)	Replaces DeepONetInterface.py. _config.py derives the DG/training/inference configs from one user JSON; deeponet_interface.py orchestrates the pipeline
dg_method/.../DGinterface.py	Explicit write_to_npz + progress_callback args, replacing the implicit resultType == "DON" caller sniffing
Data handoff	DG results → per-source HDF5 training sets, simulation parameters, and a validation split
Registration	methods-config.json + example_settings/deeponet_setting.json (frontend schema)
Packaging	requirements.txt → pyproject.toml + uv.lock; Dockerfile reworked for arm64 (gmsh via apt)

Progress is one 0–100% figure spanning DG and training. The DG half wraps TSI_TI.step_dt at runtime — TODO, as the clean fix belongs in edg-acoustics upstream.

Tests

• 30 tests, uv run pytest from deeponet_method/: 17 config builders, 6 user-facing JSON contract, 4 SimulationMethod base, 3 CLI end-to-end against a mocked stack.
• Needs gmsh for collection — run inside the container on arm64 hosts.

Current limitations

• Single source per run — CHORAS exposes one source position and DG solves one per run, so training data covers a single source.
• Therefore no speedup yet — a one-source surrogate generalizes only across receivers, which DG already gives directly. This is a proof of concept that the pipeline works end to end, not a usable surrogate.
• Multi-source input in CHORAS is the prerequisite for training to be worthwhile; the loop already iterates source_positions, so it is forward-compatible.
• No model reuse — every run retrains from scratch; no checkpoint caching or train-once/infer-many path.
• Training is inline in the simulation request, so cost is DG + full training, not inference alone.