Retether Keras to v3.11.3

[t-kalinowski]

New/Updated Activations, Losses, Optimizers

• activation_sparse_sigmoid() — LoRA-related sparse activation.
• loss_categorical_generalized_cross_entropy() — generalized CE loss wrapper.
• optimizer_muon() — new Muon optimizer (with AdamW fallback, LoRA exclusions, Newton–Schulz
  controls).

New Config functions

• Added getters/setters for training limits:
  • config_max_epochs(), config_set_max_epochs()
  • config_max_steps_per_epoch(), config_set_max_steps_per_epoch()
  • Shared config_training_limits Rd page documents both caps and their env-var overrides.
• Added config_is_nnx_enabled() to surface the JAX NNX feature flag.

Layers & Random Augmentations

• layer_random_elastic_transform() — new preprocessing layer
• layer_conv*transopose gain output_padding arg.

Expanded Ops Coverage

• Numerical/FFT/window helpers: op_angle(), op_bartlett(), op_blackman(), op_cbrt(), op_corrcoef(),
  op_deg2rad(), op_hamming(), op_hanning(), op_heaviside(), op_kaiser().
• Neural/statistical helpers: op_layer_normalization(), op_sparse_sigmoid().
• Complex utilities: op_view_as_complex(), op_view_as_real().
• Image augmentation: op_image_elastic_transform().

New S3 Methods

• Added Arg() method for both keras.src.backend.Tensor and
  keras.src.backend.common.keras_tensor.KerasTensor.