[BUGFIX] Honor Hardtanh min_val and max_val

[andrenepomuceno]

Fixes #342

Problem

The trainer exports torch.nn.Hardtanh with its bounds, {"type": "Hardtanh", "min_val": ..., "max_val": ...} (nam/models/_activations.py, ~L253–258 in neural-amp-modeler 0.13). ActivationConfig::from_json ignored both keys, and Hardtanh always clamped to [-1, 1], so a model trained with other bounds rendered differently from the trainer.

Change

This follows the existing LeakyReLU pattern: a shared singleton for the default, and a dedicated instance when parameters are given.

• ActivationConfig::from_json reads the optional min_val / max_val for Hardtanh. These are the existing ActivationConfig fields that LeakyHardtanh already uses.
• ActivationHardTanh gets a (min_val, max_val) constructor, and hard_tanh gets a matching overload. The defaults are unchanged (-1, 1).
• get_activation(config) returns a parameterized ActivationHardTanh when either bound is set (a missing bound defaults to ±1). Otherwise it returns the "Hardtanh" singleton as before.

A bare "Hardtanh" string, or an object without bounds, behaves exactly as before. No other path special-cases Hardtanh: the fast-tanh and LUT swaps only touch Tanh/Sigmoid/SiLU, and the A2 fast path only matches LeakyReLU.

Tests

• test_hardtanh_config: sets the bounds on ActivationConfig directly and checks clamping at both bounds and pass-through inside them.
• test_from_json_hardtanh_bounds: parses the trainer's export format and checks the parsed bounds and the resulting activation.
• test_from_json_hardtanh_default: "Hardtanh" and {"type": "Hardtanh"} leave the bounds unset, resolve to the singleton, and clamp to [-1, 1].
• test_hardtanh_bounds_apply_realtime_safe: applying a bounded Hardtanh allocates nothing (allocation tracking).

Cross-check: a randomly initialised neural-amp-modeler 0.13 WaveNet with inactive layer1x1 and a bounded Hardtanh, rendered with tools/render, matches the PyTorch forward pass (pad_start=True, 4096 samples) to a max abs error of 7.5e-8 with this PR and #341 (inactive layer1x1 residual) applied, down from 1.9e-3 on main.

run_tests passes in both the default and -DNAM_USE_INLINE_GEMM builds (g++ 11, Linux). Before the fix, the new tests fail.

Note

Like LeakyHardtanh, this does not check that min_val < max_val. PyTorch asserts it at construction, so a trainer-exported file always satisfies it.