[Refactor] Generalize magi_compile for non-module callables and simplify internals

[cennn]

🗂️ PR Category

• ✨ New Feature
• 🚀 Optimization (performance, memory, etc.)
• 💥 Breaking Change
• 🐛 Bug Fix
• 🛠️ Development / Refactoring
• 📚 Documentation
• 🧹 Chore (Dependencies, CI/CD, Configuration, etc.)
• 🧪 Testing

📝 Description

Non-module callable support

magi_compile now works uniformly on any callable class/instance, not just nn.Module:

• Class decoration hooks __init__ (same mechanism for Module and non-module).
• Instance decoration patches the target method on the instance directly.
• Bound method decoration compiles on __self__.
• New method_name parameter allows explicit method selection (defaults to forward).

AOT instance compilation fix

Previously, AOT mode for instance-level compilation (magi_compile(model)) had issues with dispatch_to_compiled_fwd — the old code special-cased nn.Module with isinstance checks and _magi_original_forward hacks. This is replaced by a clean two-category dispatch:

• Method path (target_method_name is set): swap target_function.__code__ and bind to instance.
• Function path: swap inspect.unwrap(obj).__code__ directly.

Both JIT and AOT use the same dispatch_to_compiled_fwd with a single dispatch_via_method flag — no more isinstance(obj, nn.Module) branches.

Internal simplification

• Remove patch.object during first JIT compilation (unnecessary — Dynamo traces compiled_entry bytecode directly, not instance.__dict__).
• Remove _state_dispatches_via_method, _magi_compile_instance (merged into _magi_compile_bound_method).
• Flatten dispatch_to_compiled_fwd: 4 sub-context-managers inlined into one.
• Remove redundant owner traversal in _mark_static_shapes (Dynamo handles self).
• Rename internal attrs for clarity: _target_callable → original_entry, _compiled_callable → compiled_entry, compiled_code → jit_compiled_code, original_code_object → original_code_for_hook.
• Centralize state attr names via get_attr_name_for_* helpers.

Tests

• Add non-module class/instance/method perf tests across all 3 suites (MLP, NormResidual, Pointwise).
• Use cuda_benchmark in API timing tests for stability.
• Lighten timing test model (dim=128×4 layers, was 256×8) for faster CI.