Restore backbone-only .compile() override on forcefield regressors

[timduignan]

Summary

Without an override, model.compile(...) falls through to nn.Module.compile() which compiles self.__call__ (= forward). That has two problems:

• DirectForcefieldRegressor: predict() calls self.model(batch) directly and never goes through __call__, so the compiled callable is created but never invoked. .compile() is silently a no-op for inference.
• ConservativeForcefieldRegressor: predict() does flow through __call__, but compiling the full regressor pulls the energy autograd backward and (for OrbMol-v2) the Coulomb / PME path into the traced graph, which dynamo fragments badly.

This PR restores the small .compile() override on both classes so self.model (the GNS message-passing backbone, where ~all the FLOPs live) is the thing actually compiled; post-backbone work runs eager. This matches what the internal core repo's orbmolv2 branch already does, and was the behaviour before the recent porting commit that exposed this regression.

Numbers

orb_v3_direct_20_omat on a single H100, 3 trials each (variance ~1%):

variant	1k atoms	10k atoms
compile=True (pre-fix default)	9.23 ms	38.36 ms
compile=False (eager)	9.23 ms	38.32 ms
manual model.model.compile(...) (this fix)	7.07 ms	30.04 ms

So the pre-fix .compile() was a literal no-op in inference for direct models, and restoring the backbone compile recovers a 22-24% inference speedup. Conservative models also benefit by getting a clean single-fragment compile graph instead of the fragmented full-regressor trace.

Tests

Adds test_compile_engages_backbone to both tests/forcefield/test_direct_regressor.py and tests/forcefield/test_conservative.py. These assert that model.model._compiled_call_impl is not None after .compile(), so this can't silently re-break if someone removes the override again in a future refactor.

Other changes

• README update: the "Full-model compilation" bullet wasn't accurate (we compile the backbone, not the full regressor). Reworded to reflect what's actually happening and the realistic speedup range.

Test plan

• pytest tests/forcefield/test_direct_regressor.py::test_compile_engages_backbone tests/forcefield/test_conservative.py::test_compile_engages_backbone passes
• Full tests/forcefield/test_{direct_regressor,conservative}.py passes in CI
• Existing test_regressor_compile_matches_eager still passes (compiled output equivalent to eager)

🤖 Generated with Claude Code

[vsimkus]

Aren't we overcorrecting here by removing the full compilation? Especially since it's passed our tests/evaluations for the conservative model. The direct regressor was indeed missing the compiled path via .predict(), but I've already put together a fix and tested it here: https://github.com/orbital-materials/orb/pull/3071

[timduignan]

Brilliant thanks yep sorry yeah this is wrong deleting