Refactor macOS export logic | Retrieve reference inputs / names / dynamic shapes from model class.

[Lewis300]

• Moves the macOS export graph contract — which inputs exist, which are state, what they're called, how they're shaped — out of export/macos.py and onto the model class as overridable hooks.
• No behavior change for anything that exports today: all 8 registered macOS models produce bit-identical reference inputs, shape bounds, and graph names.
• Fixes two real bugs found on the way: exports with a context ≤ 2048 crashed inside torch.export, and activation calibration read the cache length off a guessed input position.

The hooks

export_input_names()  -> {graph: (name, ...)}
export_state_names()  -> {graph: (name, ...)}
export_output_names() -> {graph: (name, ...)}
build_reference_inputs(config, target_dtype, spec) -> {graph: {param: tensor}}
build_dynamic_shapes(config, spec)                 -> {graph: shapes}

Keyed by graph name — a macOS model has one graph, main. The defaults are what macos.py hardcoded, so export_macos_model and the pipeline's quantization step now just call them, and validate_export_contract cross-checks that all five agree.

Two ordering rules: reference inputs bind to the traced signature, so they must be in its exact order; name lists are looked up by name, so each carries only the relative order of its own kind.

iOS classes inherit the macOS-shaped defaults, which are wrong in every respect, so they raise until the iOS follow-up implements them.

[carinapeng]

Comments about error handling, everything else looks good. Thanks!