Refactor postprocessing into a single batch-handling path; add standalone postprocessor runner

[almilder]

Summary

• Moves saving of fit results from postprocessing into the main fitter, and unifies the previously-duplicated batch-creation logic between fitter and postprocess (new build_batch/build_angular_batch/unbatch_fitted_params helpers in loops.py).
• Makes the postprocessor optional (other:run_postprocess) so a run can be restarted or replotted without redoing the fit.
• Adds run_postprocessor.py / tsadar/postprocess_runner.py, a standalone tool to replay postprocess() on an already-completed fit, either from a local run directory or a remote MLflow run id/URL.
• Deletes angular_optax, an orphaned duplicate of angular_multiple_optax's training loop with no remaining callers.
• Fixes an existing test (test_manifest.py) whose process_data mock didn't match the new signature, and a pre-existing Windows-only MLflow artifact-URI bug unrelated to this change.

Test plan

• pytest tests/ — 125 passed, 5 skipped, run on this branch alone (based on current main)
• New regression tests: tests/test_inverse/test_refit.py (guards a real historical crash in the refit path), tests/test_postprocess_runner.py (exercises both local and remote replay entry points end-to-end)

[joglekara]

Requesting changes for these correctness issues:

• Angular replay skips the ang_res_unit conversion used during fitting, so it postprocesses a different lineout range.
• refit_bad_fits passes every index to one_d_loop with num_batches=1, so it optimizes lineout 0 instead of the flagged lineout.
• Replay rebuilds checkpoints from the original config, but angular refinement mutates fe.nvx; refined checkpoints therefore cannot be deserialized.
• Remote replay also assumes defaults.yaml/inputs.yaml, excluding app runs that only have config.yaml.

CPU/GPU checks are green, but these paths need regression coverage.

[almilder]

Thanks for the thorough review — all four are fixed in the latest commit (978a6b99).

Angular replay skips the ang_res_unit conversion. Confirmed: multirun_angular_optax divides lineouts.start/end by ang_res_unit before batching during the original fit, but postprocess_runner was replaying from a freshly-loaded config that never went through that mutation. Factored the conversion out into apply_ang_res_unit() (loops.py), called from both multirun_angular_optax (unchanged behavior) and postprocess_runner.run_postprocess before it touches all_data. Also worth noting for the record: this same value feeds ThomsonScatteringDiagnostic.reduce_ATS_to_resunit's slice on the theory side too, gated on spectype == "angular_full" specifically — so the fix needs to run before LossFunction is constructed, which it now does.

refit_bad_fits always refits lineout 0. Confirmed: it passed the entire sample_indices array into one_d_loop with num_batches=1; with batch_size forced to 1, that reshapes to (N, 1) and the loop only ever touches batch index 0 — lineout 0, never the flagged lineout i. Now passes np.array([i]) so the one batch actually contains the flagged lineout.

Refined checkpoints can't be deserialized. Confirmed: multirun_angular_optax's num_mins refinement loop grows nvx (and the smoothing window length) between passes, but replay was building the ThomsonParams skeleton from the un-refined config, so eqx.tree_deserialise_leaves hit a shape mismatch whenever num_mins > 1. Factored the shape progression (not the array interpolation, which isn't needed to build a skeleton) into advance_refinement_shape(), and run_postprocess now replays it num_mins - 1 times before constructing the skeleton.

Remote replay assumes defaults.yaml/inputs.yaml. Confirmed: app-originated runs (runner.run_for_app) only ever log a single config.yaml, never defaults.yaml/inputs.yaml. Both _load_merged_config (local path) and run_postprocess_remote (remote path) now check for config.yaml first and only fall back to the defaults/inputs pair if it's absent.

Added regression coverage for all four in tests/test_inverse/test_loops.py and new cases in tests/test_postprocess_runner.py. Full suite: 133 passed, 5 skipped (up from 125).

Separately, while chasing the angular-replay fix I found that plain "angular" (non-"range" lineout type) fitting looks structurally broken independent of this PR — build_angular_batch assumes a full CCD frame that only "angular_full" actually produces, and there's no batched per-lineout path for it. Traced the dispatch pattern back to the 2024-08 introduction of the optax-based fitter and it's been this way throughout, so it predates this PR and isn't something introduced here. Leaving it out of scope for this PR — tracking separately.