Core GNPE extensions for BNS support

[stephengreen]

Summary

• Add _preprocess_data() hook to GNPESampler for subclass data preprocessing (e.g., heterodyning, decimation) before the Gibbs loop
• Add fixed_context_parameters to GNPESampler to pin proxy parameters (e.g., chirp_mass_proxy) that should not be iterated during GNPE sampling
• Add FixedInitSampler class for providing known-ahead-of-time proxy values without a trained init network
• Add return_rho2opt parameter to d_inner_h_complex_multi() for returning optimal SNR squared alongside complex inner products (needed for synthetic phase normalization)

These are the foundational changes needed for BNS single-step GNPE, where chirp mass heterodyning is implemented via the existing GNPE framework with num_iterations=1 and a fixed chirp mass proxy.

Test plan

• New tests/core/test_samplers.py with 7 tests for FixedInitSampler
• New tests/gw/test_likelihood.py with 9 tests for inner product methods and likelihood decomposition
• All existing tests pass (27 core + 2 likelihood frequency masking)

🤖 Generated with Claude Code

[stephengreen]

Note on _preprocess_data() design

On main, GNPESampler._run_sampler uses transform_pre in two distinct roles:

(1) One-time data prep before the Gibbs loop: self.init_sampler.transform_pre(context) -- borrows the init sampler's chain (whiten, decimate, repackage, to_torch).

(2) Per-iteration inside the Gibbs loop: self.transform_pre(x) -- the GWSamplerGNPE's own chain (proxy sampling, time shifts, context standardization).

Max's branch refactors this by pulling the one-time data prep into GWSamplerMixin.initialize_transforms_pre(), so GNPESampler owns its own transform_pre rather than borrowing the init sampler's. He renames the in-loop transforms to transform_gnpe_loop_pre/post. His motivation: for BNS with MFD and fixed chirp_mass_proxy, he needs to insert HeterodynePhase before decimation in the one-time chain, which requires mutating self.transform_pre.transforms.insert(0, ...) -- awkward to do on a borrowed object.

_preprocess_data() achieves the same goal more simply: the GW subclass overrides it to heterodyne raw context before calling super()._preprocess_data() (which delegates to init_sampler). This avoids the mixin refactor and the insert() mutation, and works because heterodyning always precedes everything in the one-time chain.

[stephengreen]

Follow-up on the above: considered also renaming the per-iteration transforms to transform_gnpe_loop_pre/post (as Max does), but decided against it. Since _preprocess_data already disambiguates the one-time step, the dual meaning of transform_pre on main is resolved. Two concepts is simpler than three. A broader sampler class refactoring can be considered later.